nodes/visualbruno-ComfyUI-Trellis2
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Added ReconViaGen code with new node "Sparse Generator with ReconViaGen"

Browse Source
This commit is contained in:
Bruno Fargnoli
2026-04-04 21:29:20 +02:00
parent a5242473f6
commit f791215b2a
45 changed files with 21517 additions and 1303 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+1
View File
@@ -14,6 +14,7 @@
| Date | Description |
| --- | --- |
| **2026-04-04** | Added node "Sparse Generator with ReconViaGen" |
| **2026-04-01** | Added node "Voxel to Mesh"<br>It replaces Remeshing to make watertight mesh |
| **2026-03-21** | Added node "Projection HighPoly to LowPoly"<br>Added node "Render MultiView" |
| **2026-03-17** | Added Inpainting Choice NS and TELEA |
example_workflows/Advanced.json
+546 -253
View File
@@ -1,22 +1,136 @@
{
"id": "cd6e2e00-83cc-4795-abf1-09b46428c270",
"revision": 0,
"last_node_id": 55,
"last_link_id": 117,
"last_node_id": 207,
"last_link_id": 416,
"nodes": [
{
"id": 10,
"type": "Preview3D",
"id": 69,
"type": "Trellis2LoadImageWithTransparency",
"pos": [
1739.3745174243877,
813.1451251994486
-377.72410571388025,
328.4870830026766
],
"size": [
1216.03125,
1078.3125
717.521556382955,
915.5313594325766
],
"flags": {},
"order": 6,
"order": 0,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": []
},
{
"name": "mask",
"type": "MASK",
"links": []
},
{
"name": "image_with_alpha",
"type": "IMAGE",
"links": [
241
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e7f9b30df7a09bcedf1c955e176754c73f983254",
"Node name for S&R": "Trellis2LoadImageWithTransparency",
"widget_ue_connectable": {}
},
"widgets_values": [
"Image_1024_00101_.png",
"image"
]
},
{
"id": 203,
"type": "PrimitiveInt",
"pos": [
-373.0898096413628,
163.12227474974895
],
"size": [
270,
82
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "INT",
"type": "INT",
"links": [
409
]
}
],
"title": "Target Face Number",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveInt"
},
"widgets_values": [
300000,
"fixed"
]
},
{
"id": 204,
"type": "PrimitiveString",
"pos": [
-368.2858630795249,
34.401714106564214
],
"size": [
270,
58
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "STRING",
"type": "STRING",
"links": [
410
]
}
],
"title": "Name",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveString"
},
"widgets_values": [
"Tank"
]
},
{
"id": 202,
"type": "Preview3D",
"pos": [
376.48890096546415,
331.2515281089069
],
"size": [
868.8731050199159,
920.1382602693357
],
"flags": {},
"order": 12,
"mode": 0,
"inputs": [
{
@@ -33,111 +147,37 @@
},
{
"name": "model_file",
"type": "STRING",
"type": "STRING,FILE_3D_GLB,FILE_3D_GLTF,FILE_3D_FBX,FILE_3D_OBJ,FILE_3D_STL,FILE_3D_USDZ,FILE_3D",
"widget": {
"name": "model_file"
},
"link": 22
"link": 398
}
],
"outputs": [],
"properties": {
"cnr_id": "comfy-core",
"ver": "0.4.0",
"Node name for S&R": "Preview3D",
"widget_ue_connectable": {},
"Last Time Model File": "ArmoredWarrior_00002_.glb",
"Scene Config": {
"showGrid": true,
"backgroundColor": "#282828",
"backgroundImage": "",
"backgroundRenderMode": "tiled"
},
"Camera Config": {
"cameraType": "perspective",
"fov": 35,
"state": {
"position": {
"x": 1.4101977762939673,
"y": 2.397771625867536,
"z": 8.997201229712031
},
"target": {
"x": 1.4732208254010746e-177,
"y": 2.5,
"z": 6.912872131937994e-178
},
"zoom": 1,
"cameraType": "perspective"
}
},
"Light Config": {
"intensity": 3
}
"ver": "0.18.1",
"Node name for S&R": "Preview3D"
},
"widgets_values": [
"ArmoredWarrior_00002_.glb",
"",
""
]
},
{
"id": 6,
"type": "Trellis2LoadImageWithTransparency",
"pos": [
138.06663994864374,
1231.972182577106
],
"size": [
470.78125,
439.8125
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": null
},
{
"name": "mask",
"type": "MASK",
"links": null
},
{
"name": "image_with_alpha",
"type": "IMAGE",
"links": [
116
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e7f9b30df7a09bcedf1c955e176754c73f983254",
"Node name for S&R": "Trellis2LoadImageWithTransparency",
"widget_ue_connectable": {}
},
"widgets_values": [
"Image_2048_00008_.png",
"image"
]
},
{
"id": 39,
"type": "Trellis2LoadModel",
"pos": [
243.7073730589844,
824.811803650631
404.6117778691903,
-564.6350558283436
],
"size": [
362.0625,
207.328125
301.71875,
202
],
"flags": {},
"order": 1,
"order": 3,
"mode": 0,
"inputs": [],
"outputs": [
@@ -145,7 +185,7 @@
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"links": [
87
411
]
}
],
@@ -156,79 +196,76 @@
"widget_ue_connectable": {}
},
"widgets_values": [
"TRELLIS.2-4B",
"microsoft/TRELLIS.2-4B",
"flash_attn",
"cuda",
true,
false
false,
"flex_gemm",
"flash_attn"
]
},
{
"id": 19,
"type": "Trellis2ExportMesh",
"id": 119,
"type": "Trellis2PreProcessImage",
"pos": [
1583.1015090615297,
594.4780695351682
413.7442818307099,
-185.99180300678682
],
"size": [
324,
146
281.8229166666667,
106
],
"flags": {},
"order": 5,
"order": 4,
"mode": 0,
"inputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"link": 115
"name": "image",
"type": "IMAGE",
"link": 241
}
],
"outputs": [
{
"name": "glb_path",
"type": "STRING",
"name": "image",
"type": "IMAGE",
"links": [
22
412
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e7f9b30df7a09bcedf1c955e176754c73f983254",
"Node name for S&R": "Trellis2ExportMesh",
"ver": "15854282a73cf231b81d52ada22e652f414a078b",
"Node name for S&R": "Trellis2PreProcessImage",
"widget_ue_connectable": {}
},
"widgets_values": [
"Trellis2Mesh",
"glb",
true
10,
false,
1024
]
},
{
"id": 45,
"type": "Trellis2MeshWithVoxelAdvancedGenerator",
"id": 196,
"type": "Trellis2FillHolesWithCuMesh",
"pos": [
687.5583313121791,
819.8533042235456
1148.203832543889,
-691.4362805609004
],
"size": [
495.8125,
866
312.4361328125,
58
],
"flags": {},
"order": 3,
"order": 6,
"mode": 0,
"inputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"link": 87
},
{
"name": "image",
"type": "IMAGE",
"link": 117
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 413
}
],
"outputs": [
@@ -236,41 +273,175 @@
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
113
392
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "dc66f263e832b103a81a021bf6cd53eec8ff28af",
"Node name for S&R": "Trellis2FillHolesWithCuMesh"
},
"widgets_values": [
1
]
},
{
"id": 197,
"type": "Trellis2ReconstructMeshWithQuad",
"pos": [
1143.7616665123248,
-576.3871604957823
],
"size": [
331.5878996659427,
142.11360851901668
],
"flags": {},
"order": 7,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 392
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
393
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "d4d66cfc9f1ce5b1c3b9a2e504ee4b24ffad48a5",
"Node name for S&R": "Trellis2ReconstructMeshWithQuad",
"widget_ue_connectable": {}
},
"widgets_values": [
1,
1024,
true,
true
]
},
{
"id": 198,
"type": "Trellis2SimplifyMesh",
"pos": [
1146.6254930330851,
-381.0428638307529
],
"size": [
270,
82
],
"flags": {},
"order": 8,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 393
},
{
"name": "target_face_num",
"type": "INT",
"widget": {
"name": "target_face_num"
},
"link": 409
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
394
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "5d4c1474ceb477f40bbd6d437f7da90f3d6d9229",
"Node name for S&R": "Trellis2SimplifyMesh"
},
"widgets_values": [
500000,
"Cumesh"
]
},
{
"id": 205,
"type": "Trellis2MeshWithVoxelAdvancedGenerator",
"pos": [
717.5499413662296,
-588.6883997158563
],
"size": [
413.1841796875,
702
],
"flags": {},
"order": 5,
"mode": 0,
"inputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"link": 411
},
{
"name": "image",
"type": "IMAGE",
"link": 412
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
413
]
},
{
"name": "bvh",
"type": "BVH",
"links": [
114
415
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a28d6cf93b0cf24aa9e7d3dccfd14f406227fbd0",
"Node name for S&R": "Trellis2MeshWithVoxelAdvancedGenerator",
"widget_ue_connectable": {}
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2MeshWithVoxelAdvancedGenerator"
},
"widgets_values": [
12345,
"randomize",
"fixed",
"1024_cascade",
25,
7.5,
0.01,
5,
12,
6.5,
0.2,
4,
12,
6.5,
0.2,
4,
7.5,
0.01,
3,
12,
3,
0.2,
0.01,
3,
999999,
4,
1,
32,
true,
0.1,
@@ -279,33 +450,129 @@
1,
0,
0.9,
true
true,
"euler"
]
},
{
"id": 54,
"type": "Trellis2PostProcessAndUnWrapAndRasterizer",
"id": 201,
"type": "Trellis2ExportMesh",
"pos": [
1262.5083613749216,
819.4992892507565
1597.089782460921,
-140.8020848589493
],
"size": [
419.125,
651.328125
270,
102
],
"flags": {},
"order": 4,
"order": 11,
"mode": 0,
"inputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"link": 416
},
{
"name": "filename_prefix",
"type": "STRING",
"widget": {
"name": "filename_prefix"
},
"link": 410
}
],
"outputs": [
{
"name": "glb_path",
"type": "STRING",
"links": [
398
]
},
{
"name": "relative_path",
"type": "STRING",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2ExportMesh"
},
"widgets_values": [
"3D/Trellis2",
"glb"
]
},
{
"id": 199,
"type": "Trellis2FillHolesWithMeshlib",
"pos": [
1151.769446503904,
-245.92060963339878
],
"size": [
302.4315956250001,
47.18107670916754
],
"flags": {},
"order": 9,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 113
"link": 394
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
414
]
},
{
"name": "holes_filled",
"type": "INT",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e8098f5641702e74133c2dea0b45a13a3ab19985",
"Node name for S&R": "Trellis2FillHolesWithMeshlib"
},
"widgets_values": []
},
{
"id": 207,
"type": "Trellis2UnWrapAndRasterizer",
"pos": [
1153.8343873953147,
-150.96148118668327
],
"size": [
419.15234375,
314
],
"flags": {},
"order": 10,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 414
},
{
"name": "bvh",
"type": "BVH",
"link": 114
"link": 415
}
],
"outputs": [
@@ -313,7 +580,7 @@
"name": "trimesh",
"type": "TRIMESH",
"links": [
115
416
]
},
{
@@ -329,144 +596,170 @@
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a354e8c4fead5c152d58b50eae2935cb2923cd72",
"Node name for S&R": "Trellis2PostProcessAndUnWrapAndRasterizer",
"widget_ue_connectable": {}
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2UnWrapAndRasterizer"
},
"widgets_values": [
60,
0,
1,
1,
4096,
true,
1,
0,
2000000,
"Cumesh",
true,
2048,
"OPAQUE",
"1024",
false,
true,
false,
false,
true
]
},
{
"id": 55,
"type": "Trellis2PreProcessImage",
"pos": [
310.038958285416,
1092.8065963153765
],
"size": [
281.78125,
79.328125
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [
{
"name": "image",
"type": "IMAGE",
"link": 116
}
],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": [
117
]
}
],
"properties": {
"widget_ue_connectable": {},
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a354e8c4fead5c152d58b50eae2935cb2923cd72",
"Node name for S&R": "Trellis2PreProcessImage"
},
"widgets_values": [
25
false,
"telea"
]
}
],
"links": [
[
22,
19,
0,
10,
241,
69,
2,
"STRING"
119,
0,
"IMAGE"
],
[
87,
39,
392,
196,
0,
45,
0,
"TRELLIS2PIPELINE"
],
[
113,
45,
0,
54,
197,
0,
"MESHWITHVOXEL"
],
[
114,
45,
393,
197,
0,
198,
0,
"MESHWITHVOXEL"
],
[
394,
198,
0,
199,
0,
"MESHWITHVOXEL"
],
[
398,
201,
0,
202,
2,
"STRING"
],
[
409,
203,
0,
198,
1,
54,
"INT"
],
[
410,
204,
0,
201,
1,
"STRING"
],
[
411,
39,
0,
205,
0,
"TRELLIS2PIPELINE"
],
[
412,
119,
0,
205,
1,
"IMAGE"
],
[
413,
205,
0,
196,
0,
"MESHWITHVOXEL"
],
[
414,
199,
0,
207,
0,
"MESHWITHVOXEL"
],
[
415,
205,
1,
207,
1,
"BVH"
],
[
115,
54,
416,
207,
0,
19,
201,
0,
"TRIMESH"
],
[
116,
6,
2,
55,
0,
"IMAGE"
],
[
117,
55,
0,
45,
1,
"IMAGE"
]
],
"groups": [],
"groups": [
{
"id": 1,
"title": "Configuration",
"bounding": [
-387.72410571388025,
-53.37077389343578,
737.521556382955,
1307.389216328689
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
},
{
"id": 2,
"title": "Generation",
"bounding": [
394.6117778691903,
-772.2962805609001,
1523.2043605195558,
1035.4596396805657
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
}
],
"config": {},
"extra": {
"workflowRendererVersion": "Vue",
"workflowRendererVersion": "LG",
"ue_links": [],
"ds": {
"scale": 0.520986848192445,
"scale": 0.5644739300537777,
"offset": [
91.64617544164359,
-275.12756326798984
717.24240821608,
973.4146936919584
]
},
"links_added_by_ue": [],
"frontendVersion": "1.37.11",
"frontendVersion": "1.42.8",
"VHS_latentpreview": false,
"VHS_latentpreviewrate": 0,
"VHS_MetadataImage": true,
example_workflows/Advanced_CustomSteps.json
+1249
View File
File diff suppressed because it is too large Load Diff
example_workflows/Advanced_CustomSteps_MeshOnly.json
+1118
View File
File diff suppressed because it is too large Load Diff
example_workflows/MultiViews.json
+933 -394
View File
File diff suppressed because it is too large Load Diff
example_workflows/MultiViews_MeshOnly.json
+836 -451
View File
File diff suppressed because it is too large Load Diff
example_workflows/Projection_6Views_Hy20.json
+8578
View File
File diff suppressed because it is too large Load Diff
example_workflows/ReconViaGen_MeshOnly.json
+1019
View File
File diff suppressed because it is too large Load Diff
example_workflows/RefineMesh.json
+810
View File
@@ -0,0 +1,810 @@
{
"id": "cd6e2e00-83cc-4795-abf1-09b46428c270",
"revision": 0,
"last_node_id": 233,
"last_link_id": 480,
"nodes": [
{
"id": 226,
"type": "Trellis2LoadMesh",
"pos": [
-385.0735510204828,
158.88137472879367
],
"size": [
270,
58
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"links": [
460
]
}
],
"title": "Original Mesh",
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2LoadMesh"
},
"widgets_values": [
""
]
},
{
"id": 119,
"type": "Trellis2PreProcessImage",
"pos": [
432.4981118440155,
-182.40504275448276
],
"size": [
281.8229166666667,
106
],
"flags": {},
"order": 5,
"mode": 0,
"inputs": [
{
"name": "image",
"type": "IMAGE",
"link": 241
}
],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": [
461
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "15854282a73cf231b81d52ada22e652f414a078b",
"Node name for S&R": "Trellis2PreProcessImage",
"widget_ue_connectable": {}
},
"widgets_values": [
10,
false,
1024
]
},
{
"id": 39,
"type": "Trellis2LoadModel",
"pos": [
418.8462464575081,
-492.4195525023097
],
"size": [
301.71875,
202
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"links": [
469
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "c5d66aa46bf1bbf903fd4fe9ff086ac774bb7e03",
"Node name for S&R": "Trellis2LoadModel",
"widget_ue_connectable": {}
},
"widgets_values": [
"microsoft/TRELLIS.2-4B",
"flash_attn",
"cuda",
true,
false,
"flex_gemm",
"flash_attn"
]
},
{
"id": 227,
"type": "Trellis2ReconstructMeshWithQuad",
"pos": [
1190.2002753188488,
-496.8180011267565
],
"size": [
331.5878996659427,
142.11360851901668
],
"flags": {},
"order": 8,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 470
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
472
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "d4d66cfc9f1ce5b1c3b9a2e504ee4b24ffad48a5",
"Node name for S&R": "Trellis2ReconstructMeshWithQuad",
"widget_ue_connectable": {}
},
"widgets_values": [
1,
1024,
true,
true
]
},
{
"id": 231,
"type": "Trellis2FillHolesWithCuMesh",
"pos": [
1203.103220151927,
-621.2330460050034
],
"size": [
312.4361328125,
58
],
"flags": {},
"order": 7,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 474
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
470
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "dc66f263e832b103a81a021bf6cd53eec8ff28af",
"Node name for S&R": "Trellis2FillHolesWithCuMesh"
},
"widgets_values": [
1
]
},
{
"id": 229,
"type": "Trellis2SimplifyMesh",
"pos": [
1194.8431366936325,
-292.14794928709534
],
"size": [
270,
82
],
"flags": {},
"order": 9,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 472
},
{
"name": "target_face_num",
"type": "INT",
"widget": {
"name": "target_face_num"
},
"link": 475
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
471
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "5d4c1474ceb477f40bbd6d437f7da90f3d6d9229",
"Node name for S&R": "Trellis2SimplifyMesh"
},
"widgets_values": [
500000,
"Cumesh"
]
},
{
"id": 228,
"type": "Trellis2FillHolesWithMeshlib",
"pos": [
1198.1583617137858,
-151.8665797237854
],
"size": [
249.284765625,
46
],
"flags": {},
"order": 10,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 471
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
476
]
},
{
"name": "holes_filled",
"type": "INT",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e8098f5641702e74133c2dea0b45a13a3ab19985",
"Node name for S&R": "Trellis2FillHolesWithMeshlib"
},
"widgets_values": []
},
{
"id": 232,
"type": "Trellis2UnWrapAndRasterizer",
"pos": [
1195.383895062337,
-50.23863923208595
],
"size": [
419.15234375,
314
],
"flags": {},
"order": 11,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 476
},
{
"name": "bvh",
"type": "BVH",
"link": 477
}
],
"outputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"links": [
478
]
},
{
"name": "base_color_texture",
"type": "IMAGE",
"links": null
},
{
"name": "metallic_roughness_texture",
"type": "IMAGE",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2UnWrapAndRasterizer"
},
"widgets_values": [
60,
0,
1,
1,
2048,
"OPAQUE",
false,
false,
false,
"telea"
]
},
{
"id": 233,
"type": "Trellis2ExportMesh",
"pos": [
1672.6947901006665,
-49.057670650420995
],
"size": [
270,
102
],
"flags": {},
"order": 12,
"mode": 0,
"inputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"link": 478
},
{
"name": "filename_prefix",
"type": "STRING",
"widget": {
"name": "filename_prefix"
},
"link": 479
}
],
"outputs": [
{
"name": "glb_path",
"type": "STRING",
"links": [
480
]
},
{
"name": "relative_path",
"type": "STRING",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2ExportMesh"
},
"widgets_values": [
"3D/Trellis2",
"glb"
]
},
{
"id": 202,
"type": "Preview3D",
"pos": [
369.6388214873717,
332.92876516342716
],
"size": [
868.8731050199159,
920.1382602693357
],
"flags": {},
"order": 13,
"mode": 0,
"inputs": [
{
"name": "camera_info",
"shape": 7,
"type": "LOAD3D_CAMERA",
"link": null
},
{
"name": "bg_image",
"shape": 7,
"type": "IMAGE",
"link": null
},
{
"name": "model_file",
"type": "STRING,FILE_3D_GLB,FILE_3D_GLTF,FILE_3D_FBX,FILE_3D_OBJ,FILE_3D_STL,FILE_3D_USDZ,FILE_3D",
"widget": {
"name": "model_file"
},
"link": 480
}
],
"outputs": [],
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "Preview3D"
},
"widgets_values": [
"",
""
]
},
{
"id": 69,
"type": "Trellis2LoadImageWithTransparency",
"pos": [
-383.4145219550909,
310.9862093076388
],
"size": [
717.521556382955,
915.5313594325766
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": []
},
{
"name": "mask",
"type": "MASK",
"links": []
},
{
"name": "image_with_alpha",
"type": "IMAGE",
"links": [
241
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e7f9b30df7a09bcedf1c955e176754c73f983254",
"Node name for S&R": "Trellis2LoadImageWithTransparency",
"widget_ue_connectable": {}
},
"widgets_values": [
"Image_1024_00101_.png",
"image"
]
},
{
"id": 203,
"type": "PrimitiveInt",
"pos": [
-379.85391723715725,
7.869022281792564
],
"size": [
270,
82
],
"flags": {},
"order": 3,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "INT",
"type": "INT",
"links": [
475
]
}
],
"title": "Target Face Number",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveInt"
},
"widgets_values": [
300000,
"fixed"
]
},
{
"id": 204,
"type": "PrimitiveString",
"pos": [
-388.041436029903,
-125.57562894305721
],
"size": [
270,
58
],
"flags": {},
"order": 4,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "STRING",
"type": "STRING",
"links": [
479
]
}
],
"title": "Name",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveString"
},
"widgets_values": [
"Tank"
]
},
{
"id": 225,
"type": "Trellis2MeshRefiner",
"pos": [
753.899038005419,
-389.13958558695583
],
"size": [
340.390625,
578
],
"flags": {},
"order": 6,
"mode": 0,
"inputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"link": 469
},
{
"name": "trimesh",
"type": "TRIMESH",
"link": 460
},
{
"name": "image",
"type": "IMAGE",
"link": 461
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
474
]
},
{
"name": "bvh",
"type": "BVH",
"links": [
477
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2MeshRefiner"
},
"widgets_values": [
12345,
"fixed",
1024,
12,
7.5,
0.01,
3,
12,
3,
0.01,
3,
999999,
true,
16,
0.1,
1,
0,
0.9,
true,
1,
"heun"
]
}
],
"links": [
[
241,
69,
2,
119,
0,
"IMAGE"
],
[
460,
226,
0,
225,
1,
"TRIMESH"
],
[
461,
119,
0,
225,
2,
"IMAGE"
],
[
469,
39,
0,
225,
0,
"TRELLIS2PIPELINE"
],
[
470,
231,
0,
227,
0,
"MESHWITHVOXEL"
],
[
471,
229,
0,
228,
0,
"MESHWITHVOXEL"
],
[
472,
227,
0,
229,
0,
"MESHWITHVOXEL"
],
[
474,
225,
0,
231,
0,
"MESHWITHVOXEL"
],
[
475,
203,
0,
229,
1,
"INT"
],
[
476,
228,
0,
232,
0,
"MESHWITHVOXEL"
],
[
477,
225,
1,
232,
1,
"BVH"
],
[
478,
232,
0,
233,
0,
"TRIMESH"
],
[
479,
204,
0,
233,
1,
"STRING"
],
[
480,
233,
0,
202,
2,
"STRING"
]
],
"groups": [
{
"id": 1,
"title": "Configuration",
"bounding": [
-414.5659226642583,
-213.34811694305725,
764.6855244821979,
1466.2928336195214
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
},
{
"id": 2,
"title": "Generation",
"bounding": [
374.5341225783577,
-726.2357215927766,
1820.8266085195564,
1004.7525102622308
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
}
],
"config": {},
"extra": {
"workflowRendererVersion": "LG",
"ue_links": [],
"ds": {
"scale": 0.6830134553650716,
"offset": [
111.46261093344825,
859.8232012529436
]
},
"links_added_by_ue": [],
"frontendVersion": "1.42.8",
"VHS_latentpreview": false,
"VHS_latentpreviewrate": 0,
"VHS_MetadataImage": true,
"VHS_KeepIntermediate": true
},
"version": 0.4
}
example_workflows/RefineMesh_MeshOnly.json
+776
View File
@@ -0,0 +1,776 @@
{
"id": "cd6e2e00-83cc-4795-abf1-09b46428c270",
"revision": 0,
"last_node_id": 234,
"last_link_id": 482,
"nodes": [
{
"id": 226,
"type": "Trellis2LoadMesh",
"pos": [
-385.0735510204828,
158.88137472879367
],
"size": [
270,
58
],
"flags": {},
"order": 0,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"links": [
460
]
}
],
"title": "Original Mesh",
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2LoadMesh"
},
"widgets_values": [
""
]
},
{
"id": 119,
"type": "Trellis2PreProcessImage",
"pos": [
432.4981118440155,
-182.40504275448276
],
"size": [
281.8229166666667,
106
],
"flags": {},
"order": 5,
"mode": 0,
"inputs": [
{
"name": "image",
"type": "IMAGE",
"link": 241
}
],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": [
461
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "15854282a73cf231b81d52ada22e652f414a078b",
"Node name for S&R": "Trellis2PreProcessImage",
"widget_ue_connectable": {}
},
"widgets_values": [
10,
false,
1024
]
},
{
"id": 39,
"type": "Trellis2LoadModel",
"pos": [
418.8462464575081,
-492.4195525023097
],
"size": [
301.71875,
202
],
"flags": {},
"order": 1,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"links": [
469
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "c5d66aa46bf1bbf903fd4fe9ff086ac774bb7e03",
"Node name for S&R": "Trellis2LoadModel",
"widget_ue_connectable": {}
},
"widgets_values": [
"microsoft/TRELLIS.2-4B",
"flash_attn",
"cuda",
true,
false,
"flex_gemm",
"flash_attn"
]
},
{
"id": 227,
"type": "Trellis2ReconstructMeshWithQuad",
"pos": [
1190.2002753188488,
-496.8180011267565
],
"size": [
331.5878996659427,
142.11360851901668
],
"flags": {},
"order": 8,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 470
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
472
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "d4d66cfc9f1ce5b1c3b9a2e504ee4b24ffad48a5",
"Node name for S&R": "Trellis2ReconstructMeshWithQuad",
"widget_ue_connectable": {}
},
"widgets_values": [
1,
1024,
true,
true
]
},
{
"id": 231,
"type": "Trellis2FillHolesWithCuMesh",
"pos": [
1203.103220151927,
-621.2330460050034
],
"size": [
312.4361328125,
58
],
"flags": {},
"order": 7,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 474
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
470
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "dc66f263e832b103a81a021bf6cd53eec8ff28af",
"Node name for S&R": "Trellis2FillHolesWithCuMesh"
},
"widgets_values": [
1
]
},
{
"id": 229,
"type": "Trellis2SimplifyMesh",
"pos": [
1194.8431366936325,
-292.14794928709534
],
"size": [
270,
82
],
"flags": {},
"order": 9,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 472
},
{
"name": "target_face_num",
"type": "INT",
"widget": {
"name": "target_face_num"
},
"link": 475
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
471
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "5d4c1474ceb477f40bbd6d437f7da90f3d6d9229",
"Node name for S&R": "Trellis2SimplifyMesh"
},
"widgets_values": [
500000,
"Cumesh"
]
},
{
"id": 202,
"type": "Preview3D",
"pos": [
369.6388214873717,
332.92876516342716
],
"size": [
868.8731050199159,
920.1382602693357
],
"flags": {},
"order": 13,
"mode": 0,
"inputs": [
{
"name": "camera_info",
"shape": 7,
"type": "LOAD3D_CAMERA",
"link": null
},
{
"name": "bg_image",
"shape": 7,
"type": "IMAGE",
"link": null
},
{
"name": "model_file",
"type": "STRING,FILE_3D_GLB,FILE_3D_GLTF,FILE_3D_FBX,FILE_3D_OBJ,FILE_3D_STL,FILE_3D_USDZ,FILE_3D",
"widget": {
"name": "model_file"
},
"link": 480
}
],
"outputs": [],
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "Preview3D"
},
"widgets_values": [
"",
""
]
},
{
"id": 69,
"type": "Trellis2LoadImageWithTransparency",
"pos": [
-383.4145219550909,
310.9862093076388
],
"size": [
717.521556382955,
915.5313594325766
],
"flags": {},
"order": 2,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": []
},
{
"name": "mask",
"type": "MASK",
"links": []
},
{
"name": "image_with_alpha",
"type": "IMAGE",
"links": [
241
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e7f9b30df7a09bcedf1c955e176754c73f983254",
"Node name for S&R": "Trellis2LoadImageWithTransparency",
"widget_ue_connectable": {}
},
"widgets_values": [
"Image_1024_00101_.png",
"image"
]
},
{
"id": 203,
"type": "PrimitiveInt",
"pos": [
-379.85391723715725,
7.869022281792564
],
"size": [
270,
82
],
"flags": {},
"order": 3,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "INT",
"type": "INT",
"links": [
475
]
}
],
"title": "Target Face Number",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveInt"
},
"widgets_values": [
300000,
"fixed"
]
},
{
"id": 204,
"type": "PrimitiveString",
"pos": [
-388.041436029903,
-125.57562894305721
],
"size": [
270,
58
],
"flags": {},
"order": 4,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "STRING",
"type": "STRING",
"links": [
479
]
}
],
"title": "Name",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveString"
},
"widgets_values": [
"Tank"
]
},
{
"id": 225,
"type": "Trellis2MeshRefiner",
"pos": [
753.899038005419,
-389.13958558695583
],
"size": [
340.390625,
578
],
"flags": {},
"order": 6,
"mode": 0,
"inputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"link": 469
},
{
"name": "trimesh",
"type": "TRIMESH",
"link": 460
},
{
"name": "image",
"type": "IMAGE",
"link": 461
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
474
]
},
{
"name": "bvh",
"type": "BVH",
"links": []
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2MeshRefiner"
},
"widgets_values": [
12345,
"fixed",
1024,
12,
7.5,
0.01,
3,
12,
3,
0.01,
3,
999999,
false,
16,
0.1,
1,
0,
0.9,
true,
1,
"heun"
]
},
{
"id": 228,
"type": "Trellis2FillHolesWithMeshlib",
"pos": [
1198.1583617137858,
-151.8665797237854
],
"size": [
249.284765625,
46
],
"flags": {},
"order": 10,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 471
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
481
]
},
{
"name": "holes_filled",
"type": "INT",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e8098f5641702e74133c2dea0b45a13a3ab19985",
"Node name for S&R": "Trellis2FillHolesWithMeshlib"
},
"widgets_values": []
},
{
"id": 234,
"type": "Trellis2MeshWithVoxelToTrimesh",
"pos": [
1183.7879296915492,
4.235191654770113
],
"size": [
349.41171875,
58
],
"flags": {},
"order": 11,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 481
}
],
"outputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"links": [
482
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2MeshWithVoxelToTrimesh"
},
"widgets_values": [
"90 degrees"
]
},
{
"id": 233,
"type": "Trellis2ExportMesh",
"pos": [
1205.1587971929546,
133.46673644186396
],
"size": [
270,
102
],
"flags": {},
"order": 12,
"mode": 0,
"inputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"link": 482
},
{
"name": "filename_prefix",
"type": "STRING",
"widget": {
"name": "filename_prefix"
},
"link": 479
}
],
"outputs": [
{
"name": "glb_path",
"type": "STRING",
"links": [
480
]
},
{
"name": "relative_path",
"type": "STRING",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2ExportMesh"
},
"widgets_values": [
"3D/Trellis2",
"glb"
]
}
],
"links": [
[
241,
69,
2,
119,
0,
"IMAGE"
],
[
460,
226,
0,
225,
1,
"TRIMESH"
],
[
461,
119,
0,
225,
2,
"IMAGE"
],
[
469,
39,
0,
225,
0,
"TRELLIS2PIPELINE"
],
[
470,
231,
0,
227,
0,
"MESHWITHVOXEL"
],
[
471,
229,
0,
228,
0,
"MESHWITHVOXEL"
],
[
472,
227,
0,
229,
0,
"MESHWITHVOXEL"
],
[
474,
225,
0,
231,
0,
"MESHWITHVOXEL"
],
[
475,
203,
0,
229,
1,
"INT"
],
[
479,
204,
0,
233,
1,
"STRING"
],
[
480,
233,
0,
202,
2,
"STRING"
],
[
481,
228,
0,
234,
0,
"MESHWITHVOXEL"
],
[
482,
234,
0,
233,
0,
"TRIMESH"
]
],
"groups": [
{
"id": 1,
"title": "Configuration",
"bounding": [
-414.5659226642583,
-213.34811694305725,
764.6855244821979,
1466.2928336195214
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
},
{
"id": 2,
"title": "Generation",
"bounding": [
374.5341225783577,
-726.2357215927766,
1227.3780156118423,
1005.7286067160883
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
}
],
"config": {},
"extra": {
"workflowRendererVersion": "LG",
"ue_links": [],
"ds": {
"scale": 0.6830134553650716,
"offset": [
200.2846180257333,
818.8284012529438
]
},
"links_added_by_ue": [],
"frontendVersion": "1.42.8",
"VHS_latentpreview": false,
"VHS_latentpreviewrate": 0,
"VHS_MetadataImage": true,
"VHS_KeepIntermediate": true
},
"version": 0.4
}
example_workflows/Simple.json
+485 -192
View File
@@ -1,19 +1,19 @@
{
"id": "440427ce-0c6a-4462-af51-639ed2f16dec",
"id": "cd6e2e00-83cc-4795-abf1-09b46428c270",
"revision": 0,
"last_node_id": 44,
"last_link_id": 90,
"last_node_id": 205,
"last_link_id": 413,
"nodes": [
{
"id": 6,
"id": 69,
"type": "Trellis2LoadImageWithTransparency",
"pos": [
160.23097908593746,
1093.4781077920234
-377.72410571388025,
328.4870830026766
],
"size": [
454.3125,
496.03125
717.521556382955,
915.5313594325766
],
"flags": {},
"order": 0,
@@ -23,18 +23,18 @@
{
"name": "image",
"type": "IMAGE",
"links": null
"links": []
},
{
"name": "mask",
"type": "MASK",
"links": null
"links": []
},
{
"name": "image_with_alpha",
"type": "IMAGE",
"links": [
89
241
]
}
],
@@ -45,20 +45,62 @@
"widget_ue_connectable": {}
},
"widgets_values": [
"Image_04291_.png",
"Image_1024_00101_.png",
"image"
]
},
{
"id": 119,
"type": "Trellis2PreProcessImage",
"pos": [
398.0246741688026,
149.83708875416676
],
"size": [
281.8229166666667,
106
],
"flags": {},
"order": 4,
"mode": 0,
"inputs": [
{
"name": "image",
"type": "IMAGE",
"link": 241
}
],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"links": [
391
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "15854282a73cf231b81d52ada22e652f414a078b",
"Node name for S&R": "Trellis2PreProcessImage",
"widget_ue_connectable": {}
},
"widgets_values": [
10,
false,
1024
]
},
{
"id": 39,
"type": "Trellis2LoadModel",
"pos": [
143.14495071454974,
622.610212818676
388.892170207283,
-228.80616406738957
],
"size": [
362.0625,
207.328125
301.71875,
202
],
"flags": {},
"order": 1,
@@ -69,7 +111,7 @@
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"links": [
79
390
]
}
],
@@ -80,37 +122,34 @@
"widget_ue_connectable": {}
},
"widgets_values": [
"TRELLIS.2-4B",
"microsoft/TRELLIS.2-4B",
"flash_attn",
"cuda",
true,
false
false,
"flex_gemm",
"flash_attn"
]
},
{
"id": 41,
"type": "Trellis2MeshWithVoxelGenerator",
"id": 196,
"type": "Trellis2FillHolesWithCuMesh",
"pos": [
673.7142651175388,
836.004782154444
1117.157533597802,
-362.86738879994687
],
"size": [
410.71875,
386
312.4361328125,
58
],
"flags": {},
"order": 3,
"order": 6,
"mode": 0,
"inputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"link": 79
},
{
"name": "image",
"type": "IMAGE",
"link": 90
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 396
}
],
"outputs": [
@@ -118,92 +157,192 @@
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
86
]
},
{
"name": "bvh",
"type": "BVH",
"links": [
87
392
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a28d6cf93b0cf24aa9e7d3dccfd14f406227fbd0",
"Node name for S&R": "Trellis2MeshWithVoxelGenerator",
"ver": "dc66f263e832b103a81a021bf6cd53eec8ff28af",
"Node name for S&R": "Trellis2FillHolesWithCuMesh"
},
"widgets_values": [
1
]
},
{
"id": 197,
"type": "Trellis2ReconstructMeshWithQuad",
"pos": [
1124.8153675662377,
-239.75157745064854
],
"size": [
331.5878996659427,
142.11360851901668
],
"flags": {},
"order": 7,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 392
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
393
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "d4d66cfc9f1ce5b1c3b9a2e504ee4b24ffad48a5",
"Node name for S&R": "Trellis2ReconstructMeshWithQuad",
"widget_ue_connectable": {}
},
"widgets_values": [
12345,
"randomize",
"1024_cascade",
12,
12,
12,
999999,
4,
32,
1,
1024,
true,
true
]
},
{
"id": 19,
"type": "Trellis2ExportMesh",
"id": 198,
"type": "Trellis2SimplifyMesh",
"pos": [
1480.0390859207857,
603.8997701786345
1134.132576655357,
-43.60066335397867
],
"size": [
324,
146
270,
82
],
"flags": {},
"order": 5,
"order": 8,
"mode": 0,
"inputs": [
{
"name": "trimesh",
"type": "TRIMESH",
"link": 88
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 393
},
{
"name": "target_face_num",
"type": "INT",
"widget": {
"name": "target_face_num"
},
"link": 409
}
],
"outputs": [
{
"name": "glb_path",
"type": "STRING",
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
22
394
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e7f9b30df7a09bcedf1c955e176754c73f983254",
"Node name for S&R": "Trellis2ExportMesh",
"widget_ue_connectable": {}
"ver": "5d4c1474ceb477f40bbd6d437f7da90f3d6d9229",
"Node name for S&R": "Trellis2SimplifyMesh"
},
"widgets_values": [
"Trellis2Mesh",
"glb",
true
500000,
"Cumesh"
]
},
{
"id": 10,
"type": "Preview3D",
"id": 203,
"type": "PrimitiveInt",
"pos": [
1595.8257224427462,
828.6814144924502
-373.0898096413628,
163.12227474974895
],
"size": [
752.25,
729.90625
270,
82
],
"flags": {},
"order": 6,
"order": 2,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "INT",
"type": "INT",
"links": [
409
]
}
],
"title": "Target Face Number",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveInt"
},
"widgets_values": [
300000,
"fixed"
]
},
{
"id": 204,
"type": "PrimitiveString",
"pos": [
-368.2858630795249,
34.401714106564214
],
"size": [
270,
58
],
"flags": {},
"order": 3,
"mode": 0,
"inputs": [],
"outputs": [
{
"name": "STRING",
"type": "STRING",
"links": [
410
]
}
],
"title": "Name",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.18.1",
"Node name for S&R": "PrimitiveString"
},
"widgets_values": [
"Tank"
]
},
{
"id": 202,
"type": "Preview3D",
"pos": [
376.48890096546415,
331.2515281089069
],
"size": [
868.8731050199159,
920.1382602693357
],
"flags": {},
"order": 12,
"mode": 0,
"inputs": [
{
@@ -220,77 +359,152 @@
},
{
"name": "model_file",
"type": "STRING",
"type": "STRING,FILE_3D_GLB,FILE_3D_GLTF,FILE_3D_FBX,FILE_3D_OBJ,FILE_3D_STL,FILE_3D_USDZ,FILE_3D",
"widget": {
"name": "model_file"
},
"link": 22
"link": 398
}
],
"outputs": [],
"properties": {
"cnr_id": "comfy-core",
"ver": "0.4.0",
"Node name for S&R": "Preview3D",
"widget_ue_connectable": {},
"Last Time Model File": "Home_2K_00001_.glb",
"Scene Config": {
"showGrid": true,
"backgroundColor": "#282828",
"backgroundImage": "",
"backgroundRenderMode": "tiled"
},
"Camera Config": {
"cameraType": "perspective",
"fov": 35,
"state": {
"position": {
"x": 12.514315832403488,
"y": 3.895614510886169,
"z": 9.139010516199487
},
"target": {
"x": 0,
"y": 2.1870527424637647,
"z": 0
},
"zoom": 1,
"cameraType": "perspective"
}
},
"Light Config": {
"intensity": 3
}
"ver": "0.18.1",
"Node name for S&R": "Preview3D"
},
"widgets_values": [
"Home_2K_00001_.glb",
"",
""
]
},
{
"id": 43,
"type": "Trellis2PostProcessAndUnWrapAndRasterizer",
"id": 199,
"type": "Trellis2FillHolesWithMeshlib",
"pos": [
1115.7105853634694,
836.2591783750445
1132.506716841996,
89.90828212755481
],
"size": [
419.125,
651.328125
249.284765625,
46
],
"flags": {},
"order": 4,
"order": 9,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 86
"link": 394
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
411
]
},
{
"name": "holes_filled",
"type": "INT",
"links": null
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "e8098f5641702e74133c2dea0b45a13a3ab19985",
"Node name for S&R": "Trellis2FillHolesWithMeshlib"
},
"widgets_values": []
},
{
"id": 195,
"type": "Trellis2MeshWithVoxelGenerator",
"pos": [
746.136770520365,
-107.27192057879526
],
"size": [
342.2818359375,
342
],
"flags": {},
"order": 5,
"mode": 0,
"inputs": [
{
"name": "pipeline",
"type": "TRELLIS2PIPELINE",
"link": 390
},
{
"name": "image",
"type": "IMAGE",
"link": 391
}
],
"outputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"links": [
396
]
},
{
"name": "bvh",
"type": "BVH",
"link": 87
"links": [
412
]
}
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2MeshWithVoxelGenerator"
},
"widgets_values": [
12345,
"fixed",
"1024_cascade",
12,
12,
12,
999999,
1,
32,
true,
true,
"euler"
]
},
{
"id": 205,
"type": "Trellis2UnWrapAndRasterizer",
"pos": [
1509.48883255403,
-231.18851954550428
],
"size": [
419.15234375,
314
],
"flags": {},
"order": 10,
"mode": 0,
"inputs": [
{
"name": "mesh",
"type": "MESHWITHVOXEL",
"link": 411
},
{
"name": "bvh",
"type": "BVH",
"link": 412
}
],
"outputs": [
@@ -298,7 +512,7 @@
"name": "trimesh",
"type": "TRIMESH",
"links": [
88
413
]
},
{
@@ -314,144 +528,223 @@
],
"properties": {
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a354e8c4fead5c152d58b50eae2935cb2923cd72",
"Node name for S&R": "Trellis2PostProcessAndUnWrapAndRasterizer",
"widget_ue_connectable": {}
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2UnWrapAndRasterizer"
},
"widgets_values": [
60,
0,
1,
1,
4096,
true,
1,
0,
2000000,
"Cumesh",
true,
2048,
"OPAQUE",
"1024",
false,
true,
false,
false,
true
false,
"telea"
]
},
{
"id": 44,
"type": "Trellis2PreProcessImage",
"id": 201,
"type": "Trellis2ExportMesh",
"pos": [
246.66298457453752,
916.7884835448142
1517.3851156356834,
145.96275340737535
],
"size": [
281.78125,
79.328125
270,
102
],
"flags": {},
"order": 2,
"order": 11,
"mode": 0,
"inputs": [
{
"name": "image",
"type": "IMAGE",
"link": 89
"name": "trimesh",
"type": "TRIMESH",
"link": 413
},
{
"name": "filename_prefix",
"type": "STRING",
"widget": {
"name": "filename_prefix"
},
"link": 410
}
],
"outputs": [
{
"name": "image",
"type": "IMAGE",
"name": "glb_path",
"type": "STRING",
"links": [
90
398
]
},
{
"name": "relative_path",
"type": "STRING",
"links": null
}
],
"properties": {
"widget_ue_connectable": {},
"aux_id": "visualbruno/ComfyUI-Trellis2",
"ver": "a354e8c4fead5c152d58b50eae2935cb2923cd72",
"Node name for S&R": "Trellis2PreProcessImage"
"ver": "a5242473f68fe91d5ef0b188f8658ae0da540237",
"Node name for S&R": "Trellis2ExportMesh"
},
"widgets_values": [
0
"3D/Trellis2",
"glb"
]
}
],
"links": [
[
22,
19,
0,
10,
241,
69,
2,
"STRING"
119,
0,
"IMAGE"
],
[
79,
390,
39,
0,
41,
195,
0,
"TRELLIS2PIPELINE"
],
[
86,
41,
391,
119,
0,
43,
195,
1,
"IMAGE"
],
[
392,
196,
0,
197,
0,
"MESHWITHVOXEL"
],
[
87,
41,
393,
197,
0,
198,
0,
"MESHWITHVOXEL"
],
[
394,
198,
0,
199,
0,
"MESHWITHVOXEL"
],
[
396,
195,
0,
196,
0,
"MESHWITHVOXEL"
],
[
398,
201,
0,
202,
2,
"STRING"
],
[
409,
203,
0,
198,
1,
43,
"INT"
],
[
410,
204,
0,
201,
1,
"STRING"
],
[
411,
199,
0,
205,
0,
"MESHWITHVOXEL"
],
[
412,
195,
1,
205,
1,
"BVH"
],
[
88,
43,
413,
205,
0,
19,
201,
0,
"TRIMESH"
],
[
89,
6,
2,
44,
0,
"IMAGE"
],
[
90,
44,
0,
41,
1,
"IMAGE"
]
],
"groups": [],
"groups": [
{
"id": 1,
"title": "Configuration",
"bounding": [
-387.72410571388025,
-53.37077389343578,
737.521556382955,
1307.389216328689
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
},
{
"id": 2,
"title": "Generation",
"bounding": [
378.892170207283,
-436.46738879994683,
1706.6665797550602,
727.1063315541132
],
"color": "#3f789e",
"font_size": 24,
"flags": {}
}
],
"config": {},
"extra": {
"workflowRendererVersion": "Vue",
"workflowRendererVersion": "LG",
"ue_links": [],
"ds": {
"scale": 0.6303940863128564,
"scale": 0.5644739300537773,
"offset": [
-52.07567851462492,
-327.7388708088908
683.341395587175,
735.5254552429528
]
},
"links_added_by_ue": [],
"frontendVersion": "1.37.11",
"frontendVersion": "1.42.8",
"VHS_latentpreview": false,
"VHS_latentpreviewrate": 0,
"VHS_MetadataImage": true,
nodes.py
+457 -5
View File
@@ -331,6 +331,7 @@ class Trellis2LoadModel:
"keep_models_loaded": ("BOOLEAN", {"default":True}),
"conv_backend": (["spconv","torchsparse","flex_gemm"],{"default":"flex_gemm"}),
"sparse_backend": (["xformers","flash_attn"],{"default":"flash_attn"}),
"use_reconviagen": ("BOOLEAN",{"default":False}),
},
}
@@ -340,7 +341,9 @@ class Trellis2LoadModel:
CATEGORY = "Trellis2Wrapper"
OUTPUT_NODE = True
def process(self, modelname, backend, device, low_vram, keep_models_loaded, conv_backend, sparse_backend):
def process(self, modelname, backend, device, low_vram, keep_models_loaded, conv_backend, sparse_backend, use_reconviagen):
import requests
os.environ['OPENCV_IO_ENABLE_OPENEXR'] = '1'
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" # Can save GPU memory
#os.environ["FLEX_GEMM_AUTOTUNE_CACHE_PATH"] = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'autotune_cache.json')
@@ -365,15 +368,19 @@ class Trellis2LoadModel:
local_dir=model_path,
local_dir_use_symlinks=False,
)
reconviagen_pipeline_file = os.path.join(folder_paths.models_dir,'microsoft','TRELLIS.2-4B','reconviagen_pipeline.json')
if not os.path.exists(reconviagen_pipeline_file):
source_reconviagen_pipeline_file = os.path.join(script_directory,'reconviagen_pipeline.json')
shutil.copyfile(source_reconviagen_pipeline_file,reconviagen_pipeline_file)
dinov3_model_path = os.path.join(folder_paths.models_dir,"facebook","dinov3-vitl16-pretrain-lvd1689m","model.safetensors")
if not os.path.exists(dinov3_model_path):
raise Exception("Facebook Dinov3 model not found in models/facebook/dinov3-vitl16-pretrain-lvd1689m folder")
trellis_image_large_path = os.path.join(folder_paths.models_dir,"microsoft","TRELLIS-image-large","ckpts","ss_dec_conv3d_16l8_fp16.safetensors")
if not os.path.exists(trellis_image_large_path):
print('Trellis-Image-Large ss_dec_conv3d_16l8_fp16 files not found. Trying to download the files from huggingface ...')
import requests
print('Trellis-Image-Large ss_dec_conv3d_16l8_fp16 files not found. Trying to download the files from huggingface ...')
url = "https://huggingface.co/microsoft/TRELLIS-image-large/resolve/main/ckpts/ss_dec_conv3d_16l8_fp16.json?download=true"
filename = os.path.join(folder_paths.models_dir,"microsoft","TRELLIS-image-large","ckpts","ss_dec_conv3d_16l8_fp16.json")
path = Path(filename)
@@ -398,12 +405,79 @@ class Trellis2LoadModel:
else:
raise Exception("Cannot download Trellis-Image-Large file ss_dec_conv3d_16l8_fp16.safetensors")
if use_reconviagen:
reconviagen_file = os.path.join(folder_paths.models_dir,'microsoft','TRELLIS.2-4B','ckpts','ss_vggt_cond.safetensors')
if not os.path.exists(reconviagen_file):
print('ReconViaGen file ss_vggt_cond.safetensors not found. Trying to download the files from huggingface ...')
url = "https://huggingface.co/Stable-X/trellis-vggt-v0-2/resolve/main/ckpts/ss_vggt_cond.safetensors?download=true"
filename = os.path.join(folder_paths.models_dir,"microsoft","TRELLIS.2-4B","ckpts","ss_vggt_cond.safetensors")
path = Path(filename)
path.parent.mkdir(parents=True, exist_ok=True)
response = requests.get(url)
if response.status_code == 200:
with open(filename, "wb") as f:
f.write(response.content)
print("Download ss_vggt_cond.safetensors complete!")
else:
raise Exception("Cannot download ReconViaGen file ss_vggt_cond.safetensors")
reconviagen_file = os.path.join(folder_paths.models_dir,'microsoft','TRELLIS.2-4B','ckpts','ss_vggt_cond.json')
if not os.path.exists(reconviagen_file):
print('ReconViaGen file ss_vggt_cond.json not found. Trying to download the files from huggingface ...')
url = "https://huggingface.co/Stable-X/trellis-vggt-v0-2/resolve/main/ckpts/ss_vggt_cond.json?download=true"
filename = os.path.join(folder_paths.models_dir,"microsoft","TRELLIS.2-4B","ckpts","ss_vggt_cond.json")
path = Path(filename)
path.parent.mkdir(parents=True, exist_ok=True)
response = requests.get(url)
if response.status_code == 200:
with open(filename, "wb") as f:
f.write(response.content)
print("Download ss_vggt_cond.json complete!")
else:
raise Exception("Cannot download ReconViaGen file ss_vggt_cond.json")
reconviagen_file = os.path.join(folder_paths.models_dir,'microsoft','TRELLIS.2-4B','ckpts','ss_flow_img_dit_L_16l8_fp16.safetensors')
if not os.path.exists(reconviagen_file):
print('ReconViaGen file ss_flow_img_dit_L_16l8_fp16.safetensors not found. Trying to download the files from huggingface ...')
url = "https://huggingface.co/Stable-X/trellis-vggt-v0-2/resolve/main/ckpts/ss_flow_img_dit_L_16l8_fp16.safetensors?download=true"
filename = os.path.join(folder_paths.models_dir,"microsoft","TRELLIS.2-4B","ckpts","ss_flow_img_dit_L_16l8_fp16.safetensors")
path = Path(filename)
path.parent.mkdir(parents=True, exist_ok=True)
response = requests.get(url)
if response.status_code == 200:
with open(filename, "wb") as f:
f.write(response.content)
print("Download ss_flow_img_dit_L_16l8_fp16.safetensors complete!")
else:
raise Exception("Cannot download ReconViaGen file ss_flow_img_dit_L_16l8_fp16.safetensors")
reconviagen_file = os.path.join(folder_paths.models_dir,'microsoft','TRELLIS.2-4B','ckpts','ss_flow_img_dit_L_16l8_fp16.json')
if not os.path.exists(reconviagen_file):
print('ReconViaGen file ss_flow_img_dit_L_16l8_fp16.json not found. Trying to download the files from huggingface ...')
url = "https://huggingface.co/Stable-X/trellis-vggt-v0-2/resolve/main/ckpts/ss_flow_img_dit_L_16l8_fp16.json?download=true"
filename = os.path.join(folder_paths.models_dir,"microsoft","TRELLIS.2-4B","ckpts","ss_flow_img_dit_L_16l8_fp16.json")
path = Path(filename)
path.parent.mkdir(parents=True, exist_ok=True)
response = requests.get(url)
if response.status_code == 200:
with open(filename, "wb") as f:
f.write(response.content)
print("Download ss_flow_img_dit_L_16l8_fp16.json complete!")
else:
raise Exception("Cannot download ReconViaGen file ss_flow_img_dit_L_16l8_fp16.json")
if modelname == "visualbruno/TRELLIS.2-4B-FP8":
use_fp8 = True
if use_reconviagen:
raise Exception("ReconViaGen cannot be used with TRELLIS.2-4B-FP8. Select microsoft/TRELLIS.2-4B")
else:
use_fp8 = False
pipeline = Trellis2ImageTo3DPipeline.from_pretrained(model_path, keep_models_loaded = keep_models_loaded, use_fp8=use_fp8)
pipeline = Trellis2ImageTo3DPipeline.from_pretrained(model_path, keep_models_loaded = keep_models_loaded, use_fp8=use_fp8, use_reconviagen=use_reconviagen)
pipeline.low_vram = low_vram
if device=="cuda":
@@ -4613,6 +4687,380 @@ class Trellis2VoxelToMesh:
return (mesh_copy,)
class Trellis2UnloadAllModels:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"input_1": (any,)
},
}
RETURN_TYPES = (any,)
RETURN_NAMES = ("output_1",)
FUNCTION = "process"
CATEGORY = "Trellis2Wrapper"
OUTPUT_NODE = True
def process(self, input_1):
print('Unloading all models ...')
if hasattr(mm, 'current_loaded_models'):
# Iterate backwards to safely remove items
for i in range(len(mm.current_loaded_models) - 1, -1, -1):
loaded_model = mm.current_loaded_models[i]
print(f"[AbsoluteUnload] Force-killing: {loaded_model.model.model.__class__.__name__}")
# Force VRAM unload
loaded_model.model_unload(1e32)
# Force System RAM unpinning (This is what the standard loop skipped)
if hasattr(loaded_model.model, 'partially_unload_ram'):
loaded_model.model.partially_unload_ram(1e32)
# Clear ComfyUI's intermediate cross-attention and tensor caches
if hasattr(mm, 'current_loaded_models'):
mm.current_loaded_models.clear()
import comfy.controlnet
if hasattr(comfy.controlnet, 'controlnet_loaded_models'):
comfy.controlnet.controlnet_loaded_models.clear()
mm.free_memory(memory_required = 1e30,
device = mm.get_torch_device(),
ram_required = 1e30)
print('Clearing cache ...')
mm.soft_empty_cache()
gc.collect()
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
print('Memory cleared')
return (input_1,)
class Trellis2SparseGeneratorWithReconViaGen:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"pipeline": ("TRELLIS2PIPELINE",),
"images": ("IMAGE",),
"seed": ("INT", {"default": 0, "min": 0, "max": 0x7fffffff}),
"sparse_structure_steps": ("INT",{"default":12, "min":1, "max":100},),
"sparse_structure_guidance_strength": ("FLOAT",{"default":6.50,"min":0.00,"max":99.99,"step":0.01}),
"sparse_structure_guidance_rescale": ("FLOAT",{"default":0.05,"min":0.00,"max":1.00,"step":0.01}),
"sparse_structure_rescale_t": ("FLOAT",{"default":4.00,"min":0.00,"max":9.99,"step":0.01}),
"sparse_structure_sampler": (["euler", "heun", "rk4", "rk5"], {"default": "euler"}),
"sparse_structure_resolution": ("INT", {"default":32,"min":32,"max":128,"step":4}),
"sparse_structure_guidance_interval_start": ("FLOAT",{"default":0.10,"min":0.00,"max":1.00,"step":0.01}),
"sparse_structure_guidance_interval_end": ("FLOAT",{"default":1.00,"min":0.00,"max":1.00,"step":0.01}),
},
}
RETURN_TYPES = ("COORDS", "INT", "TRELLIS2PIPELINE",)
RETURN_NAMES = ("coords", "sparse_structure_resolution", "pipeline",)
FUNCTION = "process"
CATEGORY = "Trellis2Wrapper"
OUTPUT_NODE = True
def process(self, pipeline, images, seed,
# sparse
sparse_structure_steps,
sparse_structure_guidance_strength,
sparse_structure_guidance_rescale,
sparse_structure_rescale_t,
sparse_structure_sampler,
sparse_structure_resolution,
sparse_structure_guidance_interval_start,
sparse_structure_guidance_interval_end,
):
self.seed_all(seed)
self.load_vggt_model(pipeline)
sparse_structure_guidance_interval = [sparse_structure_guidance_interval_start,sparse_structure_guidance_interval_end]
sparse_structure_sampler_params = {"steps":sparse_structure_steps,"guidance_strength":sparse_structure_guidance_strength,"guidance_rescale":sparse_structure_guidance_rescale,"guidance_interval":sparse_structure_guidance_interval,"rescale_t":sparse_structure_rescale_t}
args = pipeline._pretrained_args
sparse_sampler_prefix = pipeline.GetSamplerName(sparse_structure_sampler)
pipeline.sparse_structure_sampler = getattr(samplers, f"Flow{sparse_sampler_prefix}GuidanceIntervalSampler")(**args['sparse_structure_sampler']['args'])
pipeline.load_sparse_structure_vggt_model()
pipeline.load_sparse_structure_vggt_cond()
if images.ndim == 3:
images = images.unsqueeze(0)
coords = self._run_ss_stage_direct(pipeline, images, sparse_structure_resolution, sparse_structure_sampler_params)
if not pipeline.keep_models_loaded:
pipeline.unload_sparse_structure_vggt_model()
pipeline.unload_sparse_structure_vggt_cond()
self.unload_vggt_model(pipeline)
return (coords, sparse_structure_resolution, pipeline,)
def load_vggt_model(self, pipeline):
if pipeline.VGGT_model is None:
from .vggt.vggt.models.vggt import VGGT
pipeline.VGGT_dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] >= 8 else torch.float16
model_path = os.path.join(folder_paths.models_dir,'recongenvia')
pipeline.VGGT_model = VGGT.from_pretrained(model_path)
pipeline.VGGT_model.to('cuda')
del pipeline.VGGT_model.depth_head
del pipeline.VGGT_model.track_head
pipeline.VGGT_model.eval()
self._init_image_cond_model(pipeline)
def unload_vggt_model(self, pipeline):
del pipeline.VGGT_model
pipeline.VGGT_model = None
del pipeline.models['image_cond_model_vggt']
pipeline.models['image_cond_model_vggt'] = None
pipeline.image_cond_model_transform = None
gc.collect()
if torch.cuda.is_available():
torch.cuda.synchronize()
torch.cuda.empty_cache()
def seed_all(self, seed: int = 0):
import random
"""
Set random seeds of all components.
"""
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
@torch.no_grad()
def _run_ss_stage_direct(
self,
pipeline,
images,
target_ss_res: int,
ss_sampler_params: dict,
) -> torch.Tensor:
"""
Run only ReconViaGen's sparse structure diffusion stage to obtain coords
directly, without proceeding to the SLAT/mesh stage.
Returns:
coords : (N, 4) int tensor [batch_idx, x, y, z] in [0, target_ss_res)
"""
cuda_device = torch.device('cuda')
if pipeline.low_vram:
pipeline.VGGT_model.to(cuda_device)
with torch.no_grad():
with torch.cuda.amp.autocast(dtype=pipeline.VGGT_dtype):
aggregated_tokens_list, _ = self.vggt_feat(pipeline, images)
b, n, _, _ = aggregated_tokens_list[0].shape
image_cond = self.encode_image(pipeline, images).reshape(b, n, -1, 1024)
ss_cond = self.get_ss_cond(pipeline, image_cond[:, :, 5:], aggregated_tokens_list, 1)
ss_flow_model = pipeline.models['sparse_structure_flow_vggt_model']
sampler_params = {**pipeline.sparse_structure_sampler_params, **ss_sampler_params}
reso = ss_flow_model.resolution
ss_noise = torch.randn(1, ss_flow_model.in_channels, reso, reso, reso).to(cuda_device)
with torch.autocast('cuda', dtype=torch.float16):
ss_latent = pipeline.sparse_structure_sampler.sample(
ss_flow_model,
ss_noise,
**ss_cond,
**sampler_params,
verbose=True,
).samples
decoder = pipeline.models['sparse_structure_decoder']
decoded = decoder(ss_latent) > 0
if target_ss_res != decoded.shape[2]:
ratio = decoded.shape[2] // target_ss_res
decoded = torch.nn.functional.max_pool3d(decoded.float(), ratio, ratio, 0) > 0.5
coords = torch.argwhere(decoded)[:, [0, 2, 3, 4]].int()
if pipeline.low_vram:
pipeline.VGGT_model.to('cpu')
decoder.to('cpu')
ss_cond = pipeline._cond_cpu(ss_cond)
torch.cuda.empty_cache()
return coords
@torch.no_grad()
def _run_ss_stage(
self,
pipeline,
images,
target_ss_res: int,
ss_sampler_params: dict,
slat_sampler_params: dict,
) -> torch.Tensor:
"""
Generate a rough mesh via vggt_pipeline, then voxelise it into
surface-only coords at target_ss_res^3 for the downstream shape/tex stages.
Returns:
coords : (N, 4) int tensor [batch_idx, x, y, z] in [0, target_ss_res)
"""
vp = self.vggt_pipeline
# vp.device is dynamic (inferred from model params), so when models are on
# CPU it returns 'cpu'. Hardcode the target cuda device instead.
cuda_device = torch.device('cuda')
if self.low_vram:
self._vggt_models_to(cuda_device)
outputs, _, _ = vp.run(
image=images,
formats=["mesh"],
preprocess_image=False,
sparse_structure_sampler_params=ss_sampler_params,
slat_sampler_params=slat_sampler_params,
)
mesh_result = outputs["mesh"][0]
coords = self._mesh_to_surface_coords(mesh_result, target_ss_res, cuda_device)
if self.low_vram:
self._vggt_models_to('cpu')
torch.cuda.empty_cache()
return coords
@torch.no_grad()
def vggt_feat(self, pipeline, image):
"""
Encode the image.
Args:
image (Union[torch.Tensor, list[Image.Image]]): The image to encode
Returns:
torch.Tensor: The encoded features.
"""
if isinstance(image, torch.Tensor):
assert image.ndim == 4, "Image tensor should be batched (B, H, W, C) or (B, C, H, W)"
# ComfyUI IMAGE tensors are (B, H, W, C); convert to (B, C, H, W)
if image.shape[-1] in (3, 4):
image = image.permute(0, 3, 1, 2)
image = F.interpolate(image, 518, mode='bilinear', align_corners=False)
image = image.to(pipeline.device)
elif isinstance(image, list):
assert all(isinstance(i, Image.Image) for i in image), "Image list should be list of PIL images"
image = [i.resize((518, 518), Image.LANCZOS) for i in image]
image = [np.array(i.convert('RGB')).astype(np.float32) / 255 for i in image]
image = [torch.from_numpy(i).permute(2, 0, 1).float() for i in image]
image = torch.stack(image).to(pipeline.device)
else:
raise ValueError(f"Unsupported type of image: {type(image)}")
with torch.no_grad():
with torch.cuda.amp.autocast(dtype=pipeline.VGGT_dtype):
# Predict attributes including cameras, depth maps, and point maps.
aggregated_tokens_list, _ = pipeline.VGGT_model.aggregator(image[None])
return aggregated_tokens_list, image
def get_ss_cond(self, pipeline, image_cond: torch.Tensor, aggregated_tokens_list: list, num_samples: int) -> dict:
"""
Get the conditioning information for the model.
Args:
image (Union[torch.Tensor, list[Image.Image]]): The image prompts.
Returns:
dict: The conditioning information
"""
cond = pipeline.models['sparse_structure_vggt_cond'](aggregated_tokens_list, image_cond)
neg_cond = torch.zeros_like(cond)
return {
'cond': cond,
'neg_cond': neg_cond,
}
def get_slat_cond(self, pipeline, image_cond: torch.Tensor, aggregated_tokens_list: list, num_samples: int) -> dict:
"""
Get the conditioning information for the model.
Args:
image (Union[torch.Tensor, list[Image.Image]]): The image prompts.
Returns:
dict: The conditioning information
"""
b, n, _, _ = aggregated_tokens_list[0].shape
cond = pipeline.models['slat_vggt_cond'](aggregated_tokens_list, image_cond).reshape(b, n, -1, 1024)
cond = [c.squeeze(1) for c in cond.split(1, dim=1)]
neg_cond = [torch.zeros_like(c) for c in cond]
return {
'cond': cond,
'neg_cond': neg_cond,
}
@torch.no_grad()
def encode_image(self, pipeline, image, w_layernorm=True) -> torch.Tensor:
"""
Encode the image.
Args:
image (Union[torch.Tensor, list[Image.Image]]): The image to encode
Returns:
torch.Tensor: The encoded features.
"""
if isinstance(image, torch.Tensor):
assert image.ndim == 4, "Image tensor should be batched (B, H, W, C) or (B, C, H, W)"
# ComfyUI IMAGE tensors are (B, H, W, C); convert to (B, C, H, W)
if image.shape[-1] in (3, 4):
image = image.permute(0, 3, 1, 2)
image = F.interpolate(image, 518, mode='bilinear', align_corners=False)
image = image.to(pipeline.device)
elif isinstance(image, list):
assert all(isinstance(i, Image.Image) for i in image), "Image list should be list of PIL images"
image = [i.resize((518, 518), Image.LANCZOS) for i in image]
image = [np.array(i.convert('RGB')).astype(np.float32) / 255 for i in image]
image = [torch.from_numpy(i).permute(2, 0, 1).float() for i in image]
image = torch.stack(image).to(pipeline.device)
else:
raise ValueError(f"Unsupported type of image: {type(image)}")
image = pipeline.image_cond_model_transform(image).to(pipeline.device)
pipeline.models['image_cond_model_vggt'].to(pipeline.device)
features = pipeline.models['image_cond_model_vggt'](image, is_training=True)['x_prenorm']
if w_layernorm:
features = F.layer_norm(features, features.shape[-1:])
return features
def _init_image_cond_model(self, pipeline, name: str = "dinov2_vitl14_reg"):
"""
Initialize the image conditioning model.
"""
try:
dinov2_model = torch.hub.load(os.path.join(torch.hub.get_dir(), 'facebookresearch_dinov2_main'), name, source='local',pretrained=True)
except:
dinov2_model = torch.hub.load('facebookresearch/dinov2', name, pretrained=True)
dinov2_model.eval()
pipeline.models['image_cond_model_vggt'] = dinov2_model
transform = transforms.Compose([
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pipeline.image_cond_model_transform = transform
NODE_CLASS_MAPPINGS = {
"Trellis2LoadModel": Trellis2LoadModel,
"Trellis2MeshWithVoxelGenerator": Trellis2MeshWithVoxelGenerator,
@@ -4668,6 +5116,8 @@ NODE_CLASS_MAPPINGS = {
"Trellis2RenderMultiView": Trellis2RenderMultiView,
"Trellis2SaveImage": Trellis2SaveImage,
"Trellis2VoxelToMesh": Trellis2VoxelToMesh,
"Trellis2UnloadAllModels": Trellis2UnloadAllModels,
"Trellis2SparseGeneratorWithReconViaGen": Trellis2SparseGeneratorWithReconViaGen,
}
@@ -4726,4 +5176,6 @@ NODE_DISPLAY_NAME_MAPPINGS = {
"Trellis2RenderMultiView": "Trellis2 - Render MultiView",
"Trellis2SaveImage": "Trellis2 - Save Image",
"Trellis2VoxelToMesh": "Trellis2 - Voxel to Mesh",
"Trellis2UnloadAllModels": "Trellis2 - Unload All ComfyUI Models",
"Trellis2SparseGeneratorWithReconViaGen": "Trellis2 - Sparse Generator with ReconViaGen",
}
pyproject.toml
+1 -1
View File
@@ -1,7 +1,7 @@
[project]
name = "trellis2"
description = "ComfyUI Wrapper for Microsoft Trellis.2 - Native and Compact Structured Latents for 3D Generation"
version = "1.0.19"
version = "1.0.20"
license = {file = "LICENSE"}
# classifiers = [
# # For OS-independent nodes (works on all operating systems)
reconviagen_pipeline.json
+98
View File
@@ -0,0 +1,98 @@
{
"name": "Trellis2ImageTo3DPipeline",
"args": {
"models": {
"sparse_structure_decoder": "microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16",
"sparse_structure_flow_model": "ckpts/ss_flow_img_dit_1_3B_64_bf16",
"shape_slat_decoder": "ckpts/shape_dec_next_dc_f16c32_fp16",
"shape_slat_flow_model_512": "ckpts/slat_flow_img2shape_dit_1_3B_512_bf16",
"shape_slat_flow_model_1024": "ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16",
"tex_slat_decoder": "ckpts/tex_dec_next_dc_f16c32_fp16",
"tex_slat_flow_model_512": "ckpts/slat_flow_imgshape2tex_dit_1_3B_512_bf16",
"tex_slat_flow_model_1024": "ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16",
"sparse_structure_vggt_cond": "ckpts/ss_vggt_cond",
"slat_vggt_cond": "ckpts/slat_vggt_cond",
"sparse_structure_flow_vggt_model": "ckpts/ss_flow_img_dit_L_16l8_fp16"
},
"sparse_structure_sampler": {
"name": "FlowEulerGuidanceIntervalSampler",
"args": {
"sigma_min": 1e-5
},
"params": {
"steps": 12,
"guidance_strength": 7.5,
"guidance_rescale": 0.7,
"guidance_interval": [0.3, 1.0],
"rescale_t": 5.0
}
},
"shape_slat_sampler": {
"name": "FlowEulerGuidanceIntervalSampler",
"args": {
"sigma_min": 1e-5
},
"params": {
"steps": 12,
"guidance_strength": 7.5,
"guidance_rescale": 0.5,
"guidance_interval": [0.3, 1.0],
"rescale_t": 3.0
}
},
"shape_slat_normalization": {
"mean": [
0.781296, 0.018091, -0.495192, -0.558457, 1.060530, 0.093252, 1.518149, -0.933218,
-0.732996, 2.604095, -0.118341, -2.143904, 0.495076, -2.179512, -2.130751, -0.996944,
0.261421, -2.217463, 1.260067, -0.150213, 3.790713, 1.481266, -1.046058, -1.523667,
-0.059621, 2.220780, 1.621212, 0.877230, 0.567247, -3.175944, -3.186688, 1.578665
],
"std": [
5.972266, 4.706852, 5.445010, 5.209927, 5.320220, 4.547237, 5.020802, 5.444004,
5.226681, 5.683095, 4.831436, 5.286469, 5.652043, 5.367606, 5.525084, 4.730578,
4.805265, 5.124013, 5.530808, 5.619001, 5.103930, 5.417670, 5.269677, 5.547194,
5.634698, 5.235274, 6.110351, 5.511298, 6.237273, 4.879207, 5.347008, 5.405691
]
},
"tex_slat_sampler": {
"name": "FlowEulerGuidanceIntervalSampler",
"args": {
"sigma_min": 1e-5
},
"params": {
"steps": 12,
"guidance_strength": 1.0,
"guidance_rescale": 0.0,
"guidance_interval": [0.6, 0.9],
"rescale_t": 3.0
}
},
"tex_slat_normalization": {
"mean": [
3.501659, 2.212398, 2.226094, 0.251093, -0.026248, -0.687364, 0.439898, -0.928075,
0.029398, -0.339596, -0.869527, 1.038479, -0.972385, 0.126042, -1.129303, 0.455149,
-1.209521, 2.069067, 0.544735, 2.569128, -0.323407, 2.293000, -1.925608, -1.217717,
1.213905, 0.971588, -0.023631, 0.106750, 2.021786, 0.250524, -0.662387, -0.768862
],
"std": [
2.665652, 2.743913, 2.765121, 2.595319, 3.037293, 2.291316, 2.144656, 2.911822,
2.969419, 2.501689, 2.154811, 3.163343, 2.621215, 2.381943, 3.186697, 3.021588,
2.295916, 3.234985, 3.233086, 2.260140, 2.874801, 2.810596, 3.292720, 2.674999,
2.680878, 2.372054, 2.451546, 2.353556, 2.995195, 2.379849, 2.786195, 2.775190
]
},
"image_cond_model": {
"name": "DinoV3FeatureExtractor",
"args": {
"model_name": "facebook/dinov3-vitl16-pretrain-lvd1689m"
}
},
"rembg_model": {
"name": "BiRefNet",
"args": {
"model_name": "briaai/RMBG-2.0"
}
},
"default_pipeline_type": "1024_cascade"
}
}
trellis2/models/__init__.py
+4 -1
View File
@@ -14,7 +14,10 @@ __attributes = {
'SparseUnetVaeEncoder': 'sc_vaes.sparse_unet_vae',
'SparseUnetVaeDecoder': 'sc_vaes.sparse_unet_vae',
'FlexiDualGridVaeEncoder': 'sc_vaes.fdg_vae',
'FlexiDualGridVaeDecoder': 'sc_vaes.fdg_vae'
'FlexiDualGridVaeDecoder': 'sc_vaes.fdg_vae',
# vggt
'ModulatedMultiViewCond': 'sparse_structure_flow',
}
__submodules = []
trellis2/models/sparse_structure_flow.py
+75 -2
View File
@@ -4,8 +4,8 @@ import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from ..modules.utils import convert_module_to, manual_cast, str_to_dtype
from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock
from ..modules.utils import convert_module_to, manual_cast, str_to_dtype, convert_module_to_f16
from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock, ModulatedTransformerCrossBlock_woT
from ..modules.attention import RotaryPositionEmbedder
@@ -247,3 +247,76 @@ class SparseStructureFlowModel(nn.Module):
h = h.permute(0, 2, 1).view(h.shape[0], h.shape[2], *[self.resolution] * 3).contiguous()
return h
class ModulatedMultiViewCond(nn.Module):
"""
Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
"""
def __init__(
self,
channels: int,
ctx_channels: int,
num_heads: int,
mlp_ratio: float = 4.0,
attn_mode: Literal["full", "windowed"] = "full",
window_size: Optional[int] = None,
shift_window: Optional[Tuple[int, int, int]] = None,
use_checkpoint: bool = False,
use_rope: bool = False,
qk_rms_norm: bool = False,
qk_rms_norm_cross: bool = False,
qkv_bias: bool = True,
share_mod: bool = False,
num_init_tokens: int = 4096,
dtype: Optional[torch.dtype] = torch.float32,
use_fp16: bool = False,
):
super().__init__()
self.cond_blocks = nn.ModuleList([
ModulatedTransformerCrossBlock_woT(
channels,
ctx_channels,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
attn_mode=attn_mode,
use_checkpoint=use_checkpoint,
use_rope=use_rope,
share_mod=share_mod,
qk_rms_norm=qk_rms_norm,
qk_rms_norm_cross=qk_rms_norm_cross,
)
for _ in range(4)
])
self.use_fp16 = use_fp16
if use_fp16:
self.dtype = torch.float16
else:
self.dtype = dtype
self.multiview_cond_tokens = nn.Parameter(torch.randn(1, num_init_tokens, channels).to(dtype))
nn.init.normal_(self.multiview_cond_tokens, std=1e-6)
self.intermediate_layer_idx = [4, 11, 17, 23]
if use_fp16:
self.convert_to_fp16()
def convert_to_fp16(self) -> None:
"""
Convert the torso of the model to float16.
"""
self.use_fp16 = True
self.dtype = torch.float16
self.cond_blocks.apply(convert_module_to_f16)
self.multiview_cond_tokens = nn.Parameter(self.multiview_cond_tokens.data.to(self.dtype))
def forward(self, aggregated_tokens_list: List, image_cond: torch.Tensor):
b = aggregated_tokens_list[0].shape[0]
patch_start_idx = 5
idx = 0
cond = self.multiview_cond_tokens.repeat(b, 1, 1)
for layer_idx in self.intermediate_layer_idx:
x = aggregated_tokens_list[layer_idx][:, :, patch_start_idx:]
# x = x.reshape(b, -1, 2048) + torch.cat([image_cond.reshape(b, -1, 1024), image_cond.reshape(b, -1, 1024)],dim=-1)
x = torch.cat([x.reshape(b, -1, 2048), image_cond.reshape(b, -1, 1024)],dim=-1).to(self.dtype)
cond = self.cond_blocks[idx](cond, x)
idx = idx + 1
return cond
trellis2/modules/transformer/modulated.py
+70 -2
View File
@@ -5,7 +5,6 @@ from ..attention import MultiHeadAttention
from ..norm import LayerNorm32
from .blocks import FeedForwardNet
class ModulatedTransformerBlock(nn.Module):
"""
Transformer block (MSA + FFN) with adaptive layer norm conditioning.
@@ -162,4 +161,73 @@ class ModulatedTransformerCrossBlock(nn.Module):
return torch.utils.checkpoint.checkpoint(self._forward, x, mod, context, phases, use_reentrant=False)
else:
return self._forward(x, mod, context, phases)
class ModulatedTransformerCrossBlock_woT(nn.Module):
"""
Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
"""
def __init__(
self,
channels: int,
ctx_channels: int,
num_heads: int,
mlp_ratio: float = 4.0,
attn_mode: Literal["full", "windowed"] = "full",
window_size: Optional[int] = None,
shift_window: Optional[Tuple[int, int, int]] = None,
use_checkpoint: bool = False,
use_rope: bool = False,
qk_rms_norm: bool = False,
qk_rms_norm_cross: bool = False,
qkv_bias: bool = True,
share_mod: bool = False,
):
super().__init__()
self.use_checkpoint = use_checkpoint
self.share_mod = share_mod
self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
self.norm2 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
self.norm3 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
self.self_attn = MultiHeadAttention(
channels,
num_heads=num_heads,
type="self",
attn_mode=attn_mode,
window_size=window_size,
shift_window=shift_window,
qkv_bias=qkv_bias,
use_rope=use_rope,
qk_rms_norm=qk_rms_norm,
)
self.cross_attn = MultiHeadAttention(
channels,
ctx_channels=ctx_channels,
num_heads=num_heads,
type="cross",
attn_mode="full",
qkv_bias=qkv_bias,
qk_rms_norm=qk_rms_norm_cross,
)
self.mlp = FeedForwardNet(
channels,
mlp_ratio=mlp_ratio,
)
def _forward(self, x: torch.Tensor, context: torch.Tensor):
h = self.norm1(x)
h = self.self_attn(h)
x = x + h
h = self.norm2(x)
h = self.cross_attn(h, context)
x = x + h
h = self.norm3(x)
h = self.mlp(h)
x = x + h
return x
def forward(self, x: torch.Tensor, context: torch.Tensor):
if self.use_checkpoint:
return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
else:
return self._forward(x, context)
trellis2/pipelines/trellis2_image_to_3d.py
+46 -2
View File
@@ -26,6 +26,8 @@ import random
from comfy.utils import ProgressBar
script_directory = os.path.dirname(os.path.abspath(__file__))
def pil2tensor(image):
return torch.from_numpy(np.array(image).astype(np.float32) / 255.0)[None,]
@@ -98,6 +100,7 @@ class Trellis2ImageTo3DPipeline(Pipeline):
self.rembg_model = rembg_model
self._low_vram = low_vram
self.default_pipeline_type = default_pipeline_type
self.VGGT_model = None
self.pbr_attr_layout = {
'base_color': slice(0, 3),
'metallic': slice(3, 4),
@@ -148,14 +151,16 @@ class Trellis2ImageTo3DPipeline(Pipeline):
torch.cuda.empty_cache()
@classmethod
def from_pretrained(cls, path: str, config_file: str = "pipeline.json", keep_models_loaded = True, use_fp8 = False) -> "Trellis2ImageTo3DPipeline":
def from_pretrained(cls, path: str, config_file: str = "pipeline.json", keep_models_loaded = True, use_fp8 = False, use_reconviagen = False) -> "Trellis2ImageTo3DPipeline":
"""
Load a pretrained model.
Args:
path (str): The path to the model. Can be either local path or a Hugging Face repository.
"""
if use_fp8:
if use_reconviagen:
config_file = "reconviagen_pipeline.json"
elif use_fp8:
config_file = "pipeline_fp8.json"
pipeline = super().from_pretrained(path, config_file)
@@ -212,6 +217,45 @@ class Trellis2ImageTo3DPipeline(Pipeline):
self.models['sparse_structure_decoder'].to(self._device)
if hasattr(self.models['sparse_structure_decoder'], 'low_vram'):
self.models['sparse_structure_decoder'].low_vram = self.low_vram
def load_sparse_structure_vggt_model(self):
if self.models['sparse_structure_flow_vggt_model'] is None:
print('Loading Sparse Structure VGGT model ...')
self.models['sparse_structure_flow_vggt_model'] = models.from_pretrained(f"{self.path}/{self._pretrained_args['models']['sparse_structure_flow_vggt_model']}")
self.models['sparse_structure_flow_vggt_model'].eval()
self.models['sparse_structure_flow_vggt_model'].to(self._device)
if self.models['sparse_structure_decoder'] is None:
self.models['sparse_structure_decoder'] = models.from_pretrained(self._pretrained_args['models']['sparse_structure_decoder'])
self.models['sparse_structure_decoder'].eval()
self.models['sparse_structure_decoder'].to(self._device)
if hasattr(self.models['sparse_structure_decoder'], 'low_vram'):
self.models['sparse_structure_decoder'].low_vram = self.low_vram
def unload_sparse_structure_vggt_model(self):
if self.models['sparse_structure_flow_vggt_model']:
del self.models['sparse_structure_flow_vggt_model']
self.models['sparse_structure_flow_vggt_model'] = None
if self.models['sparse_structure_decoder']:
del self.models['sparse_structure_decoder']
self.models['sparse_structure_decoder'] = None
self._cleanup_cuda()
def load_sparse_structure_vggt_cond(self):
if self.models['sparse_structure_vggt_cond'] is None:
print('Loading Sparse Structure VGGT cond ...')
self.models['sparse_structure_vggt_cond'] = models.from_pretrained(f"{self.path}/{self._pretrained_args['models']['sparse_structure_vggt_cond']}")
self.models['sparse_structure_vggt_cond'].eval()
self.models['sparse_structure_vggt_cond'].to(self._device)
def unload_sparse_structure_vggt_cond(self):
if self.models['sparse_structure_vggt_cond']:
del self.models['sparse_structure_vggt_cond']
self.models['sparse_structure_vggt_cond'] = None
self._cleanup_cuda()
def unload_sparse_structure_model(self):
if self.models['sparse_structure_flow_model']:
vggt/vggt/heads/camera_head.py
+162
View File
@@ -0,0 +1,162 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..layers import Mlp
from ..layers.block import Block
from ..heads.head_act import activate_pose
class CameraHead(nn.Module):
"""
CameraHead predicts camera parameters from token representations using iterative refinement.
It applies a series of transformer blocks (the "trunk") to dedicated camera tokens.
"""
def __init__(
self,
dim_in: int = 2048,
trunk_depth: int = 4,
pose_encoding_type: str = "absT_quaR_FoV",
num_heads: int = 16,
mlp_ratio: int = 4,
init_values: float = 0.01,
trans_act: str = "linear",
quat_act: str = "linear",
fl_act: str = "relu", # Field of view activations: ensures FOV values are positive.
):
super().__init__()
if pose_encoding_type == "absT_quaR_FoV":
self.target_dim = 9
else:
raise ValueError(f"Unsupported camera encoding type: {pose_encoding_type}")
self.trans_act = trans_act
self.quat_act = quat_act
self.fl_act = fl_act
self.trunk_depth = trunk_depth
# Build the trunk using a sequence of transformer blocks.
self.trunk = nn.Sequential(
*[
Block(
dim=dim_in,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
init_values=init_values,
)
for _ in range(trunk_depth)
]
)
# Normalizations for camera token and trunk output.
self.token_norm = nn.LayerNorm(dim_in)
self.trunk_norm = nn.LayerNorm(dim_in)
# Learnable empty camera pose token.
self.empty_pose_tokens = nn.Parameter(torch.zeros(1, 1, self.target_dim))
self.embed_pose = nn.Linear(self.target_dim, dim_in)
# Module for producing modulation parameters: shift, scale, and a gate.
self.poseLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim_in, 3 * dim_in, bias=True))
# Adaptive layer normalization without affine parameters.
self.adaln_norm = nn.LayerNorm(dim_in, elementwise_affine=False, eps=1e-6)
self.pose_branch = Mlp(
in_features=dim_in,
hidden_features=dim_in // 2,
out_features=self.target_dim,
drop=0,
)
def forward(self, aggregated_tokens_list: list, num_iterations: int = 4) -> list:
"""
Forward pass to predict camera parameters.
Args:
aggregated_tokens_list (list): List of token tensors from the network;
the last tensor is used for prediction.
num_iterations (int, optional): Number of iterative refinement steps. Defaults to 4.
Returns:
list: A list of predicted camera encodings (post-activation) from each iteration.
"""
# Use tokens from the last block for camera prediction.
tokens = aggregated_tokens_list[-1]
# Extract the camera tokens
pose_tokens = tokens[:, :, 0]
pose_tokens = self.token_norm(pose_tokens)
pred_pose_enc_list = self.trunk_fn(pose_tokens, num_iterations)
return pred_pose_enc_list
def trunk_fn(self, pose_tokens: torch.Tensor, num_iterations: int) -> list:
"""
Iteratively refine camera pose predictions.
Args:
pose_tokens (torch.Tensor): Normalized camera tokens with shape [B, 1, C].
num_iterations (int): Number of refinement iterations.
Returns:
list: List of activated camera encodings from each iteration.
"""
B, S, C = pose_tokens.shape # S is expected to be 1.
pred_pose_enc = None
pred_pose_enc_list = []
for _ in range(num_iterations):
# Use a learned empty pose for the first iteration.
if pred_pose_enc is None:
module_input = self.embed_pose(self.empty_pose_tokens.expand(B, S, -1))
else:
# Detach the previous prediction to avoid backprop through time.
pred_pose_enc = pred_pose_enc.detach()
module_input = self.embed_pose(pred_pose_enc)
# Generate modulation parameters and split them into shift, scale, and gate components.
shift_msa, scale_msa, gate_msa = self.poseLN_modulation(module_input).chunk(3, dim=-1)
# Adaptive layer normalization and modulation.
pose_tokens_modulated = gate_msa * modulate(self.adaln_norm(pose_tokens), shift_msa, scale_msa)
pose_tokens_modulated = pose_tokens_modulated + pose_tokens
pose_tokens_modulated = self.trunk(pose_tokens_modulated)
# Compute the delta update for the pose encoding.
pred_pose_enc_delta = self.pose_branch(self.trunk_norm(pose_tokens_modulated))
if pred_pose_enc is None:
pred_pose_enc = pred_pose_enc_delta
else:
pred_pose_enc = pred_pose_enc + pred_pose_enc_delta
# Apply final activation functions for translation, quaternion, and field-of-view.
activated_pose = activate_pose(
pred_pose_enc,
trans_act=self.trans_act,
quat_act=self.quat_act,
fl_act=self.fl_act,
)
pred_pose_enc_list.append(activated_pose)
return pred_pose_enc_list
def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
"""
Modulate the input tensor using scaling and shifting parameters.
"""
# modified from https://github.com/facebookresearch/DiT/blob/796c29e532f47bba17c5b9c5eb39b9354b8b7c64/models.py#L19
return x * (1 + scale) + shift
vggt/vggt/heads/dpt_head.py
+497
View File
@@ -0,0 +1,497 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# Inspired by https://github.com/DepthAnything/Depth-Anything-V2
import os
from typing import List, Dict, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from .head_act import activate_head
from .utils import create_uv_grid, position_grid_to_embed
class DPTHead(nn.Module):
"""
DPT Head for dense prediction tasks.
This implementation follows the architecture described in "Vision Transformers for Dense Prediction"
(https://arxiv.org/abs/2103.13413). The DPT head processes features from a vision transformer
backbone and produces dense predictions by fusing multi-scale features.
Args:
dim_in (int): Input dimension (channels).
patch_size (int, optional): Patch size. Default is 14.
output_dim (int, optional): Number of output channels. Default is 4.
activation (str, optional): Activation type. Default is "inv_log".
conf_activation (str, optional): Confidence activation type. Default is "expp1".
features (int, optional): Feature channels for intermediate representations. Default is 256.
out_channels (List[int], optional): Output channels for each intermediate layer.
intermediate_layer_idx (List[int], optional): Indices of layers from aggregated tokens used for DPT.
pos_embed (bool, optional): Whether to use positional embedding. Default is True.
feature_only (bool, optional): If True, return features only without the last several layers and activation head. Default is False.
down_ratio (int, optional): Downscaling factor for the output resolution. Default is 1.
"""
def __init__(
self,
dim_in: int,
patch_size: int = 14,
output_dim: int = 4,
activation: str = "inv_log",
conf_activation: str = "expp1",
features: int = 256,
out_channels: List[int] = [256, 512, 1024, 1024],
intermediate_layer_idx: List[int] = [4, 11, 17, 23],
pos_embed: bool = True,
feature_only: bool = False,
down_ratio: int = 1,
) -> None:
super(DPTHead, self).__init__()
self.patch_size = patch_size
self.activation = activation
self.conf_activation = conf_activation
self.pos_embed = pos_embed
self.feature_only = feature_only
self.down_ratio = down_ratio
self.intermediate_layer_idx = intermediate_layer_idx
self.norm = nn.LayerNorm(dim_in)
# Projection layers for each output channel from tokens.
self.projects = nn.ModuleList(
[
nn.Conv2d(
in_channels=dim_in,
out_channels=oc,
kernel_size=1,
stride=1,
padding=0,
)
for oc in out_channels
]
)
# Resize layers for upsampling feature maps.
self.resize_layers = nn.ModuleList(
[
nn.ConvTranspose2d(
in_channels=out_channels[0], out_channels=out_channels[0], kernel_size=4, stride=4, padding=0
),
nn.ConvTranspose2d(
in_channels=out_channels[1], out_channels=out_channels[1], kernel_size=2, stride=2, padding=0
),
nn.Identity(),
nn.Conv2d(
in_channels=out_channels[3], out_channels=out_channels[3], kernel_size=3, stride=2, padding=1
),
]
)
self.scratch = _make_scratch(
out_channels,
features,
expand=False,
)
# Attach additional modules to scratch.
self.scratch.stem_transpose = None
self.scratch.refinenet1 = _make_fusion_block(features)
self.scratch.refinenet2 = _make_fusion_block(features)
self.scratch.refinenet3 = _make_fusion_block(features)
self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False)
head_features_1 = features
head_features_2 = 32
if feature_only:
self.scratch.output_conv1 = nn.Conv2d(head_features_1, head_features_1, kernel_size=3, stride=1, padding=1)
else:
self.scratch.output_conv1 = nn.Conv2d(
head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1
)
conv2_in_channels = head_features_1 // 2
self.scratch.output_conv2 = nn.Sequential(
nn.Conv2d(conv2_in_channels, head_features_2, kernel_size=3, stride=1, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0),
)
def forward(
self,
aggregated_tokens_list: List[torch.Tensor],
images: torch.Tensor,
patch_start_idx: int,
frames_chunk_size: int = 8,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
Forward pass through the DPT head, supports processing by chunking frames.
Args:
aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
images (Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
patch_start_idx (int): Starting index for patch tokens in the token sequence.
Used to separate patch tokens from other tokens (e.g., camera or register tokens).
frames_chunk_size (int, optional): Number of frames to process in each chunk.
If None or larger than S, all frames are processed at once. Default: 8.
Returns:
Tensor or Tuple[Tensor, Tensor]:
- If feature_only=True: Feature maps with shape [B, S, C, H, W]
- Otherwise: Tuple of (predictions, confidence) both with shape [B, S, 1, H, W]
"""
B, S, _, H, W = images.shape
# If frames_chunk_size is not specified or greater than S, process all frames at once
if frames_chunk_size is None or frames_chunk_size >= S:
return self._forward_impl(aggregated_tokens_list, images, patch_start_idx)
# Otherwise, process frames in chunks to manage memory usage
assert frames_chunk_size > 0
# Process frames in batches
all_preds = []
all_conf = []
for frames_start_idx in range(0, S, frames_chunk_size):
frames_end_idx = min(frames_start_idx + frames_chunk_size, S)
# Process batch of frames
if self.feature_only:
chunk_output = self._forward_impl(
aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
)
all_preds.append(chunk_output)
else:
chunk_preds, chunk_conf = self._forward_impl(
aggregated_tokens_list, images, patch_start_idx, frames_start_idx, frames_end_idx
)
all_preds.append(chunk_preds)
all_conf.append(chunk_conf)
# Concatenate results along the sequence dimension
if self.feature_only:
return torch.cat(all_preds, dim=1)
else:
return torch.cat(all_preds, dim=1), torch.cat(all_conf, dim=1)
def _forward_impl(
self,
aggregated_tokens_list: List[torch.Tensor],
images: torch.Tensor,
patch_start_idx: int,
frames_start_idx: int = None,
frames_end_idx: int = None,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
Implementation of the forward pass through the DPT head.
This method processes a specific chunk of frames from the sequence.
Args:
aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers.
images (Tensor): Input images with shape [B, S, 3, H, W].
patch_start_idx (int): Starting index for patch tokens.
frames_start_idx (int, optional): Starting index for frames to process.
frames_end_idx (int, optional): Ending index for frames to process.
Returns:
Tensor or Tuple[Tensor, Tensor]: Feature maps or (predictions, confidence).
"""
if frames_start_idx is not None and frames_end_idx is not None:
images = images[:, frames_start_idx:frames_end_idx]
B, S, _, H, W = images.shape
patch_h, patch_w = H // self.patch_size, W // self.patch_size
out = []
dpt_idx = 0
for layer_idx in self.intermediate_layer_idx:
x = aggregated_tokens_list[layer_idx][:, :, patch_start_idx:]
# Select frames if processing a chunk
if frames_start_idx is not None and frames_end_idx is not None:
x = x[:, frames_start_idx:frames_end_idx]
x = x.view(B * S, -1, x.shape[-1])
x = self.norm(x)
x = x.permute(0, 2, 1).reshape((x.shape[0], x.shape[-1], patch_h, patch_w))
x = self.projects[dpt_idx](x)
if self.pos_embed:
x = self._apply_pos_embed(x, W, H)
x = self.resize_layers[dpt_idx](x)
out.append(x)
dpt_idx += 1
# Fuse features from multiple layers.
out = self.scratch_forward(out)
# Interpolate fused output to match target image resolution.
out = custom_interpolate(
out,
(int(patch_h * self.patch_size / self.down_ratio), int(patch_w * self.patch_size / self.down_ratio)),
mode="bilinear",
align_corners=True,
)
if self.pos_embed:
out = self._apply_pos_embed(out, W, H)
if self.feature_only:
return out.view(B, S, *out.shape[1:])
out = self.scratch.output_conv2(out)
preds, conf = activate_head(out, activation=self.activation, conf_activation=self.conf_activation)
preds = preds.view(B, S, *preds.shape[1:])
conf = conf.view(B, S, *conf.shape[1:])
return preds, conf
def _apply_pos_embed(self, x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
"""
Apply positional embedding to tensor x.
"""
patch_w = x.shape[-1]
patch_h = x.shape[-2]
pos_embed = create_uv_grid(patch_w, patch_h, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
pos_embed = position_grid_to_embed(pos_embed, x.shape[1])
pos_embed = pos_embed * ratio
pos_embed = pos_embed.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1)
return x + pos_embed
def scratch_forward(self, features: List[torch.Tensor]) -> torch.Tensor:
"""
Forward pass through the fusion blocks.
Args:
features (List[Tensor]): List of feature maps from different layers.
Returns:
Tensor: Fused feature map.
"""
layer_1, layer_2, layer_3, layer_4 = features
layer_1_rn = self.scratch.layer1_rn(layer_1)
layer_2_rn = self.scratch.layer2_rn(layer_2)
layer_3_rn = self.scratch.layer3_rn(layer_3)
layer_4_rn = self.scratch.layer4_rn(layer_4)
out = self.scratch.refinenet4(layer_4_rn, size=layer_3_rn.shape[2:])
del layer_4_rn, layer_4
out = self.scratch.refinenet3(out, layer_3_rn, size=layer_2_rn.shape[2:])
del layer_3_rn, layer_3
out = self.scratch.refinenet2(out, layer_2_rn, size=layer_1_rn.shape[2:])
del layer_2_rn, layer_2
out = self.scratch.refinenet1(out, layer_1_rn)
del layer_1_rn, layer_1
out = self.scratch.output_conv1(out)
return out
################################################################################
# Modules
################################################################################
def _make_fusion_block(features: int, size: int = None, has_residual: bool = True, groups: int = 1) -> nn.Module:
return FeatureFusionBlock(
features,
nn.ReLU(inplace=True),
deconv=False,
bn=False,
expand=False,
align_corners=True,
size=size,
has_residual=has_residual,
groups=groups,
)
def _make_scratch(in_shape: List[int], out_shape: int, groups: int = 1, expand: bool = False) -> nn.Module:
scratch = nn.Module()
out_shape1 = out_shape
out_shape2 = out_shape
out_shape3 = out_shape
if len(in_shape) >= 4:
out_shape4 = out_shape
if expand:
out_shape1 = out_shape
out_shape2 = out_shape * 2
out_shape3 = out_shape * 4
if len(in_shape) >= 4:
out_shape4 = out_shape * 8
scratch.layer1_rn = nn.Conv2d(
in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
)
scratch.layer2_rn = nn.Conv2d(
in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
)
scratch.layer3_rn = nn.Conv2d(
in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
)
if len(in_shape) >= 4:
scratch.layer4_rn = nn.Conv2d(
in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
)
return scratch
class ResidualConvUnit(nn.Module):
"""Residual convolution module."""
def __init__(self, features, activation, bn, groups=1):
"""Init.
Args:
features (int): number of features
"""
super().__init__()
self.bn = bn
self.groups = groups
self.conv1 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
self.conv2 = nn.Conv2d(features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups)
self.norm1 = None
self.norm2 = None
self.activation = activation
self.skip_add = nn.quantized.FloatFunctional()
def forward(self, x):
"""Forward pass.
Args:
x (tensor): input
Returns:
tensor: output
"""
out = self.activation(x)
out = self.conv1(out)
if self.norm1 is not None:
out = self.norm1(out)
out = self.activation(out)
out = self.conv2(out)
if self.norm2 is not None:
out = self.norm2(out)
return self.skip_add.add(out, x)
class FeatureFusionBlock(nn.Module):
"""Feature fusion block."""
def __init__(
self,
features,
activation,
deconv=False,
bn=False,
expand=False,
align_corners=True,
size=None,
has_residual=True,
groups=1,
):
"""Init.
Args:
features (int): number of features
"""
super(FeatureFusionBlock, self).__init__()
self.deconv = deconv
self.align_corners = align_corners
self.groups = groups
self.expand = expand
out_features = features
if self.expand == True:
out_features = features // 2
self.out_conv = nn.Conv2d(
features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=self.groups
)
if has_residual:
self.resConfUnit1 = ResidualConvUnit(features, activation, bn, groups=self.groups)
self.has_residual = has_residual
self.resConfUnit2 = ResidualConvUnit(features, activation, bn, groups=self.groups)
self.skip_add = nn.quantized.FloatFunctional()
self.size = size
def forward(self, *xs, size=None):
"""Forward pass.
Returns:
tensor: output
"""
output = xs[0]
if self.has_residual:
res = self.resConfUnit1(xs[1])
output = self.skip_add.add(output, res)
output = self.resConfUnit2(output)
if (size is None) and (self.size is None):
modifier = {"scale_factor": 2}
elif size is None:
modifier = {"size": self.size}
else:
modifier = {"size": size}
output = custom_interpolate(output, **modifier, mode="bilinear", align_corners=self.align_corners)
output = self.out_conv(output)
return output
def custom_interpolate(
x: torch.Tensor,
size: Tuple[int, int] = None,
scale_factor: float = None,
mode: str = "bilinear",
align_corners: bool = True,
) -> torch.Tensor:
"""
Custom interpolate to avoid INT_MAX issues in nn.functional.interpolate.
"""
if size is None:
size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor))
INT_MAX = 1610612736
input_elements = size[0] * size[1] * x.shape[0] * x.shape[1]
if input_elements > INT_MAX:
chunks = torch.chunk(x, chunks=(input_elements // INT_MAX) + 1, dim=0)
interpolated_chunks = [
nn.functional.interpolate(chunk, size=size, mode=mode, align_corners=align_corners) for chunk in chunks
]
x = torch.cat(interpolated_chunks, dim=0)
return x.contiguous()
else:
return nn.functional.interpolate(x, size=size, mode=mode, align_corners=align_corners)
vggt/vggt/heads/head_act.py
+125
View File
@@ -0,0 +1,125 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn.functional as F
def activate_pose(pred_pose_enc, trans_act="linear", quat_act="linear", fl_act="linear"):
"""
Activate pose parameters with specified activation functions.
Args:
pred_pose_enc: Tensor containing encoded pose parameters [translation, quaternion, focal length]
trans_act: Activation type for translation component
quat_act: Activation type for quaternion component
fl_act: Activation type for focal length component
Returns:
Activated pose parameters tensor
"""
T = pred_pose_enc[..., :3]
quat = pred_pose_enc[..., 3:7]
fl = pred_pose_enc[..., 7:] # or fov
T = base_pose_act(T, trans_act)
quat = base_pose_act(quat, quat_act)
fl = base_pose_act(fl, fl_act) # or fov
pred_pose_enc = torch.cat([T, quat, fl], dim=-1)
return pred_pose_enc
def base_pose_act(pose_enc, act_type="linear"):
"""
Apply basic activation function to pose parameters.
Args:
pose_enc: Tensor containing encoded pose parameters
act_type: Activation type ("linear", "inv_log", "exp", "relu")
Returns:
Activated pose parameters
"""
if act_type == "linear":
return pose_enc
elif act_type == "inv_log":
return inverse_log_transform(pose_enc)
elif act_type == "exp":
return torch.exp(pose_enc)
elif act_type == "relu":
return F.relu(pose_enc)
else:
raise ValueError(f"Unknown act_type: {act_type}")
def activate_head(out, activation="norm_exp", conf_activation="expp1"):
"""
Process network output to extract 3D points and confidence values.
Args:
out: Network output tensor (B, C, H, W)
activation: Activation type for 3D points
conf_activation: Activation type for confidence values
Returns:
Tuple of (3D points tensor, confidence tensor)
"""
# Move channels from last dim to the 4th dimension => (B, H, W, C)
fmap = out.permute(0, 2, 3, 1) # B,H,W,C expected
# Split into xyz (first C-1 channels) and confidence (last channel)
xyz = fmap[:, :, :, :-1]
conf = fmap[:, :, :, -1]
if activation == "norm_exp":
d = xyz.norm(dim=-1, keepdim=True).clamp(min=1e-8)
xyz_normed = xyz / d
pts3d = xyz_normed * torch.expm1(d)
elif activation == "norm":
pts3d = xyz / xyz.norm(dim=-1, keepdim=True)
elif activation == "exp":
pts3d = torch.exp(xyz)
elif activation == "relu":
pts3d = F.relu(xyz)
elif activation == "inv_log":
pts3d = inverse_log_transform(xyz)
elif activation == "xy_inv_log":
xy, z = xyz.split([2, 1], dim=-1)
z = inverse_log_transform(z)
pts3d = torch.cat([xy * z, z], dim=-1)
elif activation == "sigmoid":
pts3d = torch.sigmoid(xyz)
elif activation == "linear":
pts3d = xyz
else:
raise ValueError(f"Unknown activation: {activation}")
if conf_activation == "expp1":
conf_out = 1 + conf.exp()
elif conf_activation == "expp0":
conf_out = conf.exp()
elif conf_activation == "sigmoid":
conf_out = torch.sigmoid(conf)
else:
raise ValueError(f"Unknown conf_activation: {conf_activation}")
return pts3d, conf_out
def inverse_log_transform(y):
"""
Apply inverse log transform: sign(y) * (exp(|y|) - 1)
Args:
y: Input tensor
Returns:
Transformed tensor
"""
return torch.sign(y) * (torch.expm1(torch.abs(y)))
vggt/vggt/heads/track_head.py
+108
View File
@@ -0,0 +1,108 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch.nn as nn
from .dpt_head import DPTHead
from .track_modules.base_track_predictor import BaseTrackerPredictor
class TrackHead(nn.Module):
"""
Track head that uses DPT head to process tokens and BaseTrackerPredictor for tracking.
The tracking is performed iteratively, refining predictions over multiple iterations.
"""
def __init__(
self,
dim_in,
patch_size=14,
features=128,
iters=4,
predict_conf=True,
stride=2,
corr_levels=7,
corr_radius=4,
hidden_size=384,
):
"""
Initialize the TrackHead module.
Args:
dim_in (int): Input dimension of tokens from the backbone.
patch_size (int): Size of image patches used in the vision transformer.
features (int): Number of feature channels in the feature extractor output.
iters (int): Number of refinement iterations for tracking predictions.
predict_conf (bool): Whether to predict confidence scores for tracked points.
stride (int): Stride value for the tracker predictor.
corr_levels (int): Number of correlation pyramid levels
corr_radius (int): Radius for correlation computation, controlling the search area.
hidden_size (int): Size of hidden layers in the tracker network.
"""
super().__init__()
self.patch_size = patch_size
# Feature extractor based on DPT architecture
# Processes tokens into feature maps for tracking
self.feature_extractor = DPTHead(
dim_in=dim_in,
patch_size=patch_size,
features=features,
feature_only=True, # Only output features, no activation
down_ratio=2, # Reduces spatial dimensions by factor of 2
pos_embed=False,
)
# Tracker module that predicts point trajectories
# Takes feature maps and predicts coordinates and visibility
self.tracker = BaseTrackerPredictor(
latent_dim=features, # Match the output_dim of feature extractor
predict_conf=predict_conf,
stride=stride,
corr_levels=corr_levels,
corr_radius=corr_radius,
hidden_size=hidden_size,
)
self.iters = iters
def forward(self, aggregated_tokens_list, images, patch_start_idx, query_points=None, iters=None):
"""
Forward pass of the TrackHead.
Args:
aggregated_tokens_list (list): List of aggregated tokens from the backbone.
images (torch.Tensor): Input images of shape (B, S, C, H, W) where:
B = batch size, S = sequence length.
patch_start_idx (int): Starting index for patch tokens.
query_points (torch.Tensor, optional): Initial query points to track.
If None, points are initialized by the tracker.
iters (int, optional): Number of refinement iterations. If None, uses self.iters.
Returns:
tuple:
- coord_preds (torch.Tensor): Predicted coordinates for tracked points.
- vis_scores (torch.Tensor): Visibility scores for tracked points.
- conf_scores (torch.Tensor): Confidence scores for tracked points (if predict_conf=True).
"""
B, S, _, H, W = images.shape
# Extract features from tokens
# feature_maps has shape (B, S, C, H//2, W//2) due to down_ratio=2
feature_maps = self.feature_extractor(aggregated_tokens_list, images, patch_start_idx)
# Use default iterations if not specified
if iters is None:
iters = self.iters
# Perform tracking using the extracted features
coord_preds, vis_scores, conf_scores = self.tracker(
query_points=query_points,
fmaps=feature_maps,
iters=iters,
)
return coord_preds, vis_scores, conf_scores
vggt/vggt/heads/track_modules/__init__.py
+5
View File
@@ -0,0 +1,5 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
vggt/vggt/heads/track_modules/base_track_predictor.py
+209
View File
@@ -0,0 +1,209 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn as nn
from einops import rearrange, repeat
from .blocks import EfficientUpdateFormer, CorrBlock
from .utils import sample_features4d, get_2d_embedding, get_2d_sincos_pos_embed
from .modules import Mlp
class BaseTrackerPredictor(nn.Module):
def __init__(
self,
stride=1,
corr_levels=5,
corr_radius=4,
latent_dim=128,
hidden_size=384,
use_spaceatt=True,
depth=6,
max_scale=518,
predict_conf=True,
):
super(BaseTrackerPredictor, self).__init__()
"""
The base template to create a track predictor
Modified from https://github.com/facebookresearch/co-tracker/
and https://github.com/facebookresearch/vggsfm
"""
self.stride = stride
self.latent_dim = latent_dim
self.corr_levels = corr_levels
self.corr_radius = corr_radius
self.hidden_size = hidden_size
self.max_scale = max_scale
self.predict_conf = predict_conf
self.flows_emb_dim = latent_dim // 2
self.corr_mlp = Mlp(
in_features=self.corr_levels * (self.corr_radius * 2 + 1) ** 2,
hidden_features=self.hidden_size,
out_features=self.latent_dim,
)
self.transformer_dim = self.latent_dim + self.latent_dim + self.latent_dim + 4
self.query_ref_token = nn.Parameter(torch.randn(1, 2, self.transformer_dim))
space_depth = depth if use_spaceatt else 0
time_depth = depth
self.updateformer = EfficientUpdateFormer(
space_depth=space_depth,
time_depth=time_depth,
input_dim=self.transformer_dim,
hidden_size=self.hidden_size,
output_dim=self.latent_dim + 2,
mlp_ratio=4.0,
add_space_attn=use_spaceatt,
)
self.fmap_norm = nn.LayerNorm(self.latent_dim)
self.ffeat_norm = nn.GroupNorm(1, self.latent_dim)
# A linear layer to update track feats at each iteration
self.ffeat_updater = nn.Sequential(nn.Linear(self.latent_dim, self.latent_dim), nn.GELU())
self.vis_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
if predict_conf:
self.conf_predictor = nn.Sequential(nn.Linear(self.latent_dim, 1))
def forward(self, query_points, fmaps=None, iters=6, return_feat=False, down_ratio=1, apply_sigmoid=True):
"""
query_points: B x N x 2, the number of batches, tracks, and xy
fmaps: B x S x C x HH x WW, the number of batches, frames, and feature dimension.
note HH and WW is the size of feature maps instead of original images
"""
B, N, D = query_points.shape
B, S, C, HH, WW = fmaps.shape
assert D == 2, "Input points must be 2D coordinates"
# apply a layernorm to fmaps here
fmaps = self.fmap_norm(fmaps.permute(0, 1, 3, 4, 2))
fmaps = fmaps.permute(0, 1, 4, 2, 3)
# Scale the input query_points because we may downsample the images
# by down_ratio or self.stride
# e.g., if a 3x1024x1024 image is processed to a 128x256x256 feature map
# its query_points should be query_points/4
if down_ratio > 1:
query_points = query_points / float(down_ratio)
query_points = query_points / float(self.stride)
# Init with coords as the query points
# It means the search will start from the position of query points at the reference frames
coords = query_points.clone().reshape(B, 1, N, 2).repeat(1, S, 1, 1)
# Sample/extract the features of the query points in the query frame
query_track_feat = sample_features4d(fmaps[:, 0], coords[:, 0])
# init track feats by query feats
track_feats = query_track_feat.unsqueeze(1).repeat(1, S, 1, 1) # B, S, N, C
# back up the init coords
coords_backup = coords.clone()
fcorr_fn = CorrBlock(fmaps, num_levels=self.corr_levels, radius=self.corr_radius)
coord_preds = []
# Iterative Refinement
for _ in range(iters):
# Detach the gradients from the last iteration
# (in my experience, not very important for performance)
coords = coords.detach()
fcorrs = fcorr_fn.corr_sample(track_feats, coords)
corr_dim = fcorrs.shape[3]
fcorrs_ = fcorrs.permute(0, 2, 1, 3).reshape(B * N, S, corr_dim)
fcorrs_ = self.corr_mlp(fcorrs_)
# Movement of current coords relative to query points
flows = (coords - coords[:, 0:1]).permute(0, 2, 1, 3).reshape(B * N, S, 2)
flows_emb = get_2d_embedding(flows, self.flows_emb_dim, cat_coords=False)
# (In my trials, it is also okay to just add the flows_emb instead of concat)
flows_emb = torch.cat([flows_emb, flows / self.max_scale, flows / self.max_scale], dim=-1)
track_feats_ = track_feats.permute(0, 2, 1, 3).reshape(B * N, S, self.latent_dim)
# Concatenate them as the input for the transformers
transformer_input = torch.cat([flows_emb, fcorrs_, track_feats_], dim=2)
# 2D positional embed
# TODO: this can be much simplified
pos_embed = get_2d_sincos_pos_embed(self.transformer_dim, grid_size=(HH, WW)).to(query_points.device)
sampled_pos_emb = sample_features4d(pos_embed.expand(B, -1, -1, -1), coords[:, 0])
sampled_pos_emb = rearrange(sampled_pos_emb, "b n c -> (b n) c").unsqueeze(1)
x = transformer_input + sampled_pos_emb
# Add the query ref token to the track feats
query_ref_token = torch.cat(
[self.query_ref_token[:, 0:1], self.query_ref_token[:, 1:2].expand(-1, S - 1, -1)], dim=1
)
x = x + query_ref_token.to(x.device).to(x.dtype)
# B, N, S, C
x = rearrange(x, "(b n) s d -> b n s d", b=B)
# Compute the delta coordinates and delta track features
delta, _ = self.updateformer(x)
# BN, S, C
delta = rearrange(delta, " b n s d -> (b n) s d", b=B)
delta_coords_ = delta[:, :, :2]
delta_feats_ = delta[:, :, 2:]
track_feats_ = track_feats_.reshape(B * N * S, self.latent_dim)
delta_feats_ = delta_feats_.reshape(B * N * S, self.latent_dim)
# Update the track features
track_feats_ = self.ffeat_updater(self.ffeat_norm(delta_feats_)) + track_feats_
track_feats = track_feats_.reshape(B, N, S, self.latent_dim).permute(0, 2, 1, 3) # BxSxNxC
# B x S x N x 2
coords = coords + delta_coords_.reshape(B, N, S, 2).permute(0, 2, 1, 3)
# Force coord0 as query
# because we assume the query points should not be changed
coords[:, 0] = coords_backup[:, 0]
# The predicted tracks are in the original image scale
if down_ratio > 1:
coord_preds.append(coords * self.stride * down_ratio)
else:
coord_preds.append(coords * self.stride)
# B, S, N
vis_e = self.vis_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
if apply_sigmoid:
vis_e = torch.sigmoid(vis_e)
if self.predict_conf:
conf_e = self.conf_predictor(track_feats.reshape(B * S * N, self.latent_dim)).reshape(B, S, N)
if apply_sigmoid:
conf_e = torch.sigmoid(conf_e)
else:
conf_e = None
if return_feat:
return coord_preds, vis_e, track_feats, query_track_feat, conf_e
else:
return coord_preds, vis_e, conf_e
vggt/vggt/heads/track_modules/blocks.py
+246
View File
@@ -0,0 +1,246 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# Modified from https://github.com/facebookresearch/co-tracker/
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from .utils import bilinear_sampler
from .modules import Mlp, AttnBlock, CrossAttnBlock, ResidualBlock
class EfficientUpdateFormer(nn.Module):
"""
Transformer model that updates track estimates.
"""
def __init__(
self,
space_depth=6,
time_depth=6,
input_dim=320,
hidden_size=384,
num_heads=8,
output_dim=130,
mlp_ratio=4.0,
add_space_attn=True,
num_virtual_tracks=64,
):
super().__init__()
self.out_channels = 2
self.num_heads = num_heads
self.hidden_size = hidden_size
self.add_space_attn = add_space_attn
# Add input LayerNorm before linear projection
self.input_norm = nn.LayerNorm(input_dim)
self.input_transform = torch.nn.Linear(input_dim, hidden_size, bias=True)
# Add output LayerNorm before final projection
self.output_norm = nn.LayerNorm(hidden_size)
self.flow_head = torch.nn.Linear(hidden_size, output_dim, bias=True)
self.num_virtual_tracks = num_virtual_tracks
if self.add_space_attn:
self.virual_tracks = nn.Parameter(torch.randn(1, num_virtual_tracks, 1, hidden_size))
else:
self.virual_tracks = None
self.time_blocks = nn.ModuleList(
[
AttnBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
attn_class=nn.MultiheadAttention,
)
for _ in range(time_depth)
]
)
if add_space_attn:
self.space_virtual_blocks = nn.ModuleList(
[
AttnBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
attn_class=nn.MultiheadAttention,
)
for _ in range(space_depth)
]
)
self.space_point2virtual_blocks = nn.ModuleList(
[CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
)
self.space_virtual2point_blocks = nn.ModuleList(
[CrossAttnBlock(hidden_size, hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(space_depth)]
)
assert len(self.time_blocks) >= len(self.space_virtual2point_blocks)
self.initialize_weights()
def initialize_weights(self):
def _basic_init(module):
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
torch.nn.init.trunc_normal_(self.flow_head.weight, std=0.001)
self.apply(_basic_init)
def forward(self, input_tensor, mask=None):
# Apply input LayerNorm
input_tensor = self.input_norm(input_tensor)
tokens = self.input_transform(input_tensor)
init_tokens = tokens
B, _, T, _ = tokens.shape
if self.add_space_attn:
virtual_tokens = self.virual_tracks.repeat(B, 1, T, 1)
tokens = torch.cat([tokens, virtual_tokens], dim=1)
_, N, _, _ = tokens.shape
j = 0
for i in range(len(self.time_blocks)):
time_tokens = tokens.contiguous().view(B * N, T, -1) # B N T C -> (B N) T C
time_tokens = self.time_blocks[i](time_tokens)
tokens = time_tokens.view(B, N, T, -1) # (B N) T C -> B N T C
if self.add_space_attn and (i % (len(self.time_blocks) // len(self.space_virtual_blocks)) == 0):
space_tokens = tokens.permute(0, 2, 1, 3).contiguous().view(B * T, N, -1) # B N T C -> (B T) N C
point_tokens = space_tokens[:, : N - self.num_virtual_tracks]
virtual_tokens = space_tokens[:, N - self.num_virtual_tracks :]
virtual_tokens = self.space_virtual2point_blocks[j](virtual_tokens, point_tokens, mask=mask)
virtual_tokens = self.space_virtual_blocks[j](virtual_tokens)
point_tokens = self.space_point2virtual_blocks[j](point_tokens, virtual_tokens, mask=mask)
space_tokens = torch.cat([point_tokens, virtual_tokens], dim=1)
tokens = space_tokens.view(B, T, N, -1).permute(0, 2, 1, 3) # (B T) N C -> B N T C
j += 1
if self.add_space_attn:
tokens = tokens[:, : N - self.num_virtual_tracks]
tokens = tokens + init_tokens
# Apply output LayerNorm before final projection
tokens = self.output_norm(tokens)
flow = self.flow_head(tokens)
return flow, None
class CorrBlock:
def __init__(self, fmaps, num_levels=4, radius=4, multiple_track_feats=False, padding_mode="zeros"):
"""
Build a pyramid of feature maps from the input.
fmaps: Tensor (B, S, C, H, W)
num_levels: number of pyramid levels (each downsampled by factor 2)
radius: search radius for sampling correlation
multiple_track_feats: if True, split the target features per pyramid level
padding_mode: passed to grid_sample / bilinear_sampler
"""
B, S, C, H, W = fmaps.shape
self.S, self.C, self.H, self.W = S, C, H, W
self.num_levels = num_levels
self.radius = radius
self.padding_mode = padding_mode
self.multiple_track_feats = multiple_track_feats
# Build pyramid: each level is half the spatial resolution of the previous
self.fmaps_pyramid = [fmaps] # level 0 is full resolution
current_fmaps = fmaps
for i in range(num_levels - 1):
B, S, C, H, W = current_fmaps.shape
# Merge batch & sequence dimensions
current_fmaps = current_fmaps.reshape(B * S, C, H, W)
# Avg pool down by factor 2
current_fmaps = F.avg_pool2d(current_fmaps, kernel_size=2, stride=2)
_, _, H_new, W_new = current_fmaps.shape
current_fmaps = current_fmaps.reshape(B, S, C, H_new, W_new)
self.fmaps_pyramid.append(current_fmaps)
# Precompute a delta grid (of shape (2r+1, 2r+1, 2)) for sampling.
# This grid is added to the (scaled) coordinate centroids.
r = self.radius
dx = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
dy = torch.linspace(-r, r, 2 * r + 1, device=fmaps.device, dtype=fmaps.dtype)
# delta: for every (dy,dx) displacement (i.e. Δx, Δy)
self.delta = torch.stack(torch.meshgrid(dy, dx, indexing="ij"), dim=-1) # shape: (2r+1, 2r+1, 2)
def corr_sample(self, targets, coords):
"""
Instead of storing the entire correlation pyramid, we compute each level's correlation
volume, sample it immediately, then discard it. This saves GPU memory.
Args:
targets: Tensor (B, S, N, C) — features for the current targets.
coords: Tensor (B, S, N, 2) — coordinates at full resolution.
Returns:
Tensor (B, S, N, L) where L = num_levels * (2*radius+1)**2 (concatenated sampled correlations)
"""
B, S, N, C = targets.shape
# If you have multiple track features, split them per level.
if self.multiple_track_feats:
targets_split = torch.split(targets, C // self.num_levels, dim=-1)
out_pyramid = []
for i, fmaps in enumerate(self.fmaps_pyramid):
# Get current spatial resolution H, W for this pyramid level.
B, S, C, H, W = fmaps.shape
# Reshape feature maps for correlation computation:
# fmap2s: (B, S, C, H*W)
fmap2s = fmaps.view(B, S, C, H * W)
# Choose appropriate target features.
fmap1 = targets_split[i] if self.multiple_track_feats else targets # shape: (B, S, N, C)
# Compute correlation directly
corrs = compute_corr_level(fmap1, fmap2s, C)
corrs = corrs.view(B, S, N, H, W)
# Prepare sampling grid:
# Scale down the coordinates for the current level.
centroid_lvl = coords.reshape(B * S * N, 1, 1, 2) / (2**i)
# Make sure our precomputed delta grid is on the same device/dtype.
delta_lvl = self.delta.to(coords.device).to(coords.dtype)
# Now the grid for grid_sample is:
# coords_lvl = centroid_lvl + delta_lvl (broadcasted over grid)
coords_lvl = centroid_lvl + delta_lvl.view(1, 2 * self.radius + 1, 2 * self.radius + 1, 2)
# Sample from the correlation volume using bilinear interpolation.
# We reshape corrs to (B * S * N, 1, H, W) so grid_sample acts over each target.
corrs_sampled = bilinear_sampler(
corrs.reshape(B * S * N, 1, H, W), coords_lvl, padding_mode=self.padding_mode
)
# The sampled output is (B * S * N, 1, 2r+1, 2r+1). Flatten the last two dims.
corrs_sampled = corrs_sampled.view(B, S, N, -1) # Now shape: (B, S, N, (2r+1)^2)
out_pyramid.append(corrs_sampled)
# Concatenate all levels along the last dimension.
out = torch.cat(out_pyramid, dim=-1).contiguous()
return out
def compute_corr_level(fmap1, fmap2s, C):
# fmap1: (B, S, N, C)
# fmap2s: (B, S, C, H*W)
corrs = torch.matmul(fmap1, fmap2s) # (B, S, N, H*W)
corrs = corrs.view(fmap1.shape[0], fmap1.shape[1], fmap1.shape[2], -1) # (B, S, N, H*W)
return corrs / math.sqrt(C)
vggt/vggt/heads/track_modules/modules.py
+218
View File
@@ -0,0 +1,218 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn as nn
import torch.nn.functional as F
from functools import partial
from typing import Callable
import collections
from torch import Tensor
from itertools import repeat
# From PyTorch internals
def _ntuple(n):
def parse(x):
if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
return tuple(x)
return tuple(repeat(x, n))
return parse
def exists(val):
return val is not None
def default(val, d):
return val if exists(val) else d
to_2tuple = _ntuple(2)
class ResidualBlock(nn.Module):
"""
ResidualBlock: construct a block of two conv layers with residual connections
"""
def __init__(self, in_planes, planes, norm_fn="group", stride=1, kernel_size=3):
super(ResidualBlock, self).__init__()
self.conv1 = nn.Conv2d(
in_planes,
planes,
kernel_size=kernel_size,
padding=1,
stride=stride,
padding_mode="zeros",
)
self.conv2 = nn.Conv2d(
planes,
planes,
kernel_size=kernel_size,
padding=1,
padding_mode="zeros",
)
self.relu = nn.ReLU(inplace=True)
num_groups = planes // 8
if norm_fn == "group":
self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
if not stride == 1:
self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
elif norm_fn == "batch":
self.norm1 = nn.BatchNorm2d(planes)
self.norm2 = nn.BatchNorm2d(planes)
if not stride == 1:
self.norm3 = nn.BatchNorm2d(planes)
elif norm_fn == "instance":
self.norm1 = nn.InstanceNorm2d(planes)
self.norm2 = nn.InstanceNorm2d(planes)
if not stride == 1:
self.norm3 = nn.InstanceNorm2d(planes)
elif norm_fn == "none":
self.norm1 = nn.Sequential()
self.norm2 = nn.Sequential()
if not stride == 1:
self.norm3 = nn.Sequential()
else:
raise NotImplementedError
if stride == 1:
self.downsample = None
else:
self.downsample = nn.Sequential(
nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride),
self.norm3,
)
def forward(self, x):
y = x
y = self.relu(self.norm1(self.conv1(y)))
y = self.relu(self.norm2(self.conv2(y)))
if self.downsample is not None:
x = self.downsample(x)
return self.relu(x + y)
class Mlp(nn.Module):
"""MLP as used in Vision Transformer, MLP-Mixer and related networks"""
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer=nn.GELU,
norm_layer=None,
bias=True,
drop=0.0,
use_conv=False,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
bias = to_2tuple(bias)
drop_probs = to_2tuple(drop)
linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
self.act = act_layer()
self.drop1 = nn.Dropout(drop_probs[0])
self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
self.drop2 = nn.Dropout(drop_probs[1])
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop1(x)
x = self.fc2(x)
x = self.drop2(x)
return x
class AttnBlock(nn.Module):
def __init__(
self,
hidden_size,
num_heads,
attn_class: Callable[..., nn.Module] = nn.MultiheadAttention,
mlp_ratio=4.0,
**block_kwargs
):
"""
Self attention block
"""
super().__init__()
self.norm1 = nn.LayerNorm(hidden_size)
self.norm2 = nn.LayerNorm(hidden_size)
self.attn = attn_class(embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs)
mlp_hidden_dim = int(hidden_size * mlp_ratio)
self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
def forward(self, x, mask=None):
# Prepare the mask for PyTorch's attention (it expects a different format)
# attn_mask = mask if mask is not None else None
# Normalize before attention
x = self.norm1(x)
# PyTorch's MultiheadAttention returns attn_output, attn_output_weights
# attn_output, _ = self.attn(x, x, x, attn_mask=attn_mask)
attn_output, _ = self.attn(x, x, x)
# Add & Norm
x = x + attn_output
x = x + self.mlp(self.norm2(x))
return x
class CrossAttnBlock(nn.Module):
def __init__(self, hidden_size, context_dim, num_heads=1, mlp_ratio=4.0, **block_kwargs):
"""
Cross attention block
"""
super().__init__()
self.norm1 = nn.LayerNorm(hidden_size)
self.norm_context = nn.LayerNorm(hidden_size)
self.norm2 = nn.LayerNorm(hidden_size)
self.cross_attn = nn.MultiheadAttention(
embed_dim=hidden_size, num_heads=num_heads, batch_first=True, **block_kwargs
)
mlp_hidden_dim = int(hidden_size * mlp_ratio)
self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, drop=0)
def forward(self, x, context, mask=None):
# Normalize inputs
x = self.norm1(x)
context = self.norm_context(context)
# Apply cross attention
# Note: nn.MultiheadAttention returns attn_output, attn_output_weights
attn_output, _ = self.cross_attn(x, context, context, attn_mask=mask)
# Add & Norm
x = x + attn_output
x = x + self.mlp(self.norm2(x))
return x
vggt/vggt/heads/track_modules/utils.py
+226
View File
@@ -0,0 +1,226 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# Modified from https://github.com/facebookresearch/vggsfm
# and https://github.com/facebookresearch/co-tracker/tree/main
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple, Union
def get_2d_sincos_pos_embed(embed_dim: int, grid_size: Union[int, Tuple[int, int]], return_grid=False) -> torch.Tensor:
"""
This function initializes a grid and generates a 2D positional embedding using sine and cosine functions.
It is a wrapper of get_2d_sincos_pos_embed_from_grid.
Args:
- embed_dim: The embedding dimension.
- grid_size: The grid size.
Returns:
- pos_embed: The generated 2D positional embedding.
"""
if isinstance(grid_size, tuple):
grid_size_h, grid_size_w = grid_size
else:
grid_size_h = grid_size_w = grid_size
grid_h = torch.arange(grid_size_h, dtype=torch.float)
grid_w = torch.arange(grid_size_w, dtype=torch.float)
grid = torch.meshgrid(grid_w, grid_h, indexing="xy")
grid = torch.stack(grid, dim=0)
grid = grid.reshape([2, 1, grid_size_h, grid_size_w])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if return_grid:
return (
pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2),
grid,
)
return pos_embed.reshape(1, grid_size_h, grid_size_w, -1).permute(0, 3, 1, 2)
def get_2d_sincos_pos_embed_from_grid(embed_dim: int, grid: torch.Tensor) -> torch.Tensor:
"""
This function generates a 2D positional embedding from a given grid using sine and cosine functions.
Args:
- embed_dim: The embedding dimension.
- grid: The grid to generate the embedding from.
Returns:
- emb: The generated 2D positional embedding.
"""
assert embed_dim % 2 == 0
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
emb = torch.cat([emb_h, emb_w], dim=2) # (H*W, D)
return emb
def get_1d_sincos_pos_embed_from_grid(embed_dim: int, pos: torch.Tensor) -> torch.Tensor:
"""
This function generates a 1D positional embedding from a given grid using sine and cosine functions.
Args:
- embed_dim: The embedding dimension.
- pos: The position to generate the embedding from.
Returns:
- emb: The generated 1D positional embedding.
"""
assert embed_dim % 2 == 0
omega = torch.arange(embed_dim // 2, dtype=torch.double)
omega /= embed_dim / 2.0
omega = 1.0 / 10000**omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = torch.einsum("m,d->md", pos, omega) # (M, D/2), outer product
emb_sin = torch.sin(out) # (M, D/2)
emb_cos = torch.cos(out) # (M, D/2)
emb = torch.cat([emb_sin, emb_cos], dim=1) # (M, D)
return emb[None].float()
def get_2d_embedding(xy: torch.Tensor, C: int, cat_coords: bool = True) -> torch.Tensor:
"""
This function generates a 2D positional embedding from given coordinates using sine and cosine functions.
Args:
- xy: The coordinates to generate the embedding from.
- C: The size of the embedding.
- cat_coords: A flag to indicate whether to concatenate the original coordinates to the embedding.
Returns:
- pe: The generated 2D positional embedding.
"""
B, N, D = xy.shape
assert D == 2
x = xy[:, :, 0:1]
y = xy[:, :, 1:2]
div_term = (torch.arange(0, C, 2, device=xy.device, dtype=torch.float32) * (1000.0 / C)).reshape(1, 1, int(C / 2))
pe_x = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
pe_y = torch.zeros(B, N, C, device=xy.device, dtype=torch.float32)
pe_x[:, :, 0::2] = torch.sin(x * div_term)
pe_x[:, :, 1::2] = torch.cos(x * div_term)
pe_y[:, :, 0::2] = torch.sin(y * div_term)
pe_y[:, :, 1::2] = torch.cos(y * div_term)
pe = torch.cat([pe_x, pe_y], dim=2) # (B, N, C*3)
if cat_coords:
pe = torch.cat([xy, pe], dim=2) # (B, N, C*3+3)
return pe
def bilinear_sampler(input, coords, align_corners=True, padding_mode="border"):
r"""Sample a tensor using bilinear interpolation
`bilinear_sampler(input, coords)` samples a tensor :attr:`input` at
coordinates :attr:`coords` using bilinear interpolation. It is the same
as `torch.nn.functional.grid_sample()` but with a different coordinate
convention.
The input tensor is assumed to be of shape :math:`(B, C, H, W)`, where
:math:`B` is the batch size, :math:`C` is the number of channels,
:math:`H` is the height of the image, and :math:`W` is the width of the
image. The tensor :attr:`coords` of shape :math:`(B, H_o, W_o, 2)` is
interpreted as an array of 2D point coordinates :math:`(x_i,y_i)`.
Alternatively, the input tensor can be of size :math:`(B, C, T, H, W)`,
in which case sample points are triplets :math:`(t_i,x_i,y_i)`. Note
that in this case the order of the components is slightly different
from `grid_sample()`, which would expect :math:`(x_i,y_i,t_i)`.
If `align_corners` is `True`, the coordinate :math:`x` is assumed to be
in the range :math:`[0,W-1]`, with 0 corresponding to the center of the
left-most image pixel :math:`W-1` to the center of the right-most
pixel.
If `align_corners` is `False`, the coordinate :math:`x` is assumed to
be in the range :math:`[0,W]`, with 0 corresponding to the left edge of
the left-most pixel :math:`W` to the right edge of the right-most
pixel.
Similar conventions apply to the :math:`y` for the range
:math:`[0,H-1]` and :math:`[0,H]` and to :math:`t` for the range
:math:`[0,T-1]` and :math:`[0,T]`.
Args:
input (Tensor): batch of input images.
coords (Tensor): batch of coordinates.
align_corners (bool, optional): Coordinate convention. Defaults to `True`.
padding_mode (str, optional): Padding mode. Defaults to `"border"`.
Returns:
Tensor: sampled points.
"""
coords = coords.detach().clone()
############################################################
# IMPORTANT:
coords = coords.to(input.device).to(input.dtype)
############################################################
sizes = input.shape[2:]
assert len(sizes) in [2, 3]
if len(sizes) == 3:
# t x y -> x y t to match dimensions T H W in grid_sample
coords = coords[..., [1, 2, 0]]
if align_corners:
scale = torch.tensor(
[2 / max(size - 1, 1) for size in reversed(sizes)], device=coords.device, dtype=coords.dtype
)
else:
scale = torch.tensor([2 / size for size in reversed(sizes)], device=coords.device, dtype=coords.dtype)
coords.mul_(scale) # coords = coords * scale
coords.sub_(1) # coords = coords - 1
return F.grid_sample(input, coords, align_corners=align_corners, padding_mode=padding_mode)
def sample_features4d(input, coords):
r"""Sample spatial features
`sample_features4d(input, coords)` samples the spatial features
:attr:`input` represented by a 4D tensor :math:`(B, C, H, W)`.
The field is sampled at coordinates :attr:`coords` using bilinear
interpolation. :attr:`coords` is assumed to be of shape :math:`(B, R,
2)`, where each sample has the format :math:`(x_i, y_i)`. This uses the
same convention as :func:`bilinear_sampler` with `align_corners=True`.
The output tensor has one feature per point, and has shape :math:`(B,
R, C)`.
Args:
input (Tensor): spatial features.
coords (Tensor): points.
Returns:
Tensor: sampled features.
"""
B, _, _, _ = input.shape
# B R 2 -> B R 1 2
coords = coords.unsqueeze(2)
# B C R 1
feats = bilinear_sampler(input, coords)
return feats.permute(0, 2, 1, 3).view(B, -1, feats.shape[1] * feats.shape[3]) # B C R 1 -> B R C
vggt/vggt/heads/utils.py
+108
View File
@@ -0,0 +1,108 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn as nn
def position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100) -> torch.Tensor:
"""
Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC)
Args:
pos_grid: Tensor of shape (H, W, 2) containing 2D coordinates
embed_dim: Output channel dimension for embeddings
Returns:
Tensor of shape (H, W, embed_dim) with positional embeddings
"""
H, W, grid_dim = pos_grid.shape
assert grid_dim == 2
pos_flat = pos_grid.reshape(-1, grid_dim) # Flatten to (H*W, 2)
# Process x and y coordinates separately
emb_x = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0) # [1, H*W, D/2]
emb_y = make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0) # [1, H*W, D/2]
# Combine and reshape
emb = torch.cat([emb_x, emb_y], dim=-1) # [1, H*W, D]
return emb.view(H, W, embed_dim) # [H, W, D]
def make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100) -> torch.Tensor:
"""
This function generates a 1D positional embedding from a given grid using sine and cosine functions.
Args:
- embed_dim: The embedding dimension.
- pos: The position to generate the embedding from.
Returns:
- emb: The generated 1D positional embedding.
"""
assert embed_dim % 2 == 0
omega = torch.arange(embed_dim // 2, dtype=torch.double, device=pos.device)
omega /= embed_dim / 2.0
omega = 1.0 / omega_0**omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = torch.einsum("m,d->md", pos, omega) # (M, D/2), outer product
emb_sin = torch.sin(out) # (M, D/2)
emb_cos = torch.cos(out) # (M, D/2)
emb = torch.cat([emb_sin, emb_cos], dim=1) # (M, D)
return emb.float()
# Inspired by https://github.com/microsoft/moge
def create_uv_grid(
width: int, height: int, aspect_ratio: float = None, dtype: torch.dtype = None, device: torch.device = None
) -> torch.Tensor:
"""
Create a normalized UV grid of shape (width, height, 2).
The grid spans horizontally and vertically according to an aspect ratio,
ensuring the top-left corner is at (-x_span, -y_span) and the bottom-right
corner is at (x_span, y_span), normalized by the diagonal of the plane.
Args:
width (int): Number of points horizontally.
height (int): Number of points vertically.
aspect_ratio (float, optional): Width-to-height ratio. Defaults to width/height.
dtype (torch.dtype, optional): Data type of the resulting tensor.
device (torch.device, optional): Device on which the tensor is created.
Returns:
torch.Tensor: A (width, height, 2) tensor of UV coordinates.
"""
# Derive aspect ratio if not explicitly provided
if aspect_ratio is None:
aspect_ratio = float(width) / float(height)
# Compute normalized spans for X and Y
diag_factor = (aspect_ratio**2 + 1.0) ** 0.5
span_x = aspect_ratio / diag_factor
span_y = 1.0 / diag_factor
# Establish the linspace boundaries
left_x = -span_x * (width - 1) / width
right_x = span_x * (width - 1) / width
top_y = -span_y * (height - 1) / height
bottom_y = span_y * (height - 1) / height
# Generate 1D coordinates
x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device)
y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device)
# Create 2D meshgrid (width x height) and stack into UV
uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy")
uv_grid = torch.stack((uu, vv), dim=-1)
return uv_grid
vggt/vggt/layers/__init__.py
+11
View File
@@ -0,0 +1,11 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from .mlp import Mlp
from .patch_embed import PatchEmbed
from .swiglu_ffn import SwiGLUFFN, SwiGLUFFNFused
from .block import NestedTensorBlock
from .attention import MemEffAttention
vggt/vggt/layers/attention.py
+98
View File
@@ -0,0 +1,98 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py
import logging
import os
import warnings
from torch import Tensor
from torch import nn
import torch.nn.functional as F
XFORMERS_AVAILABLE = False
class Attention(nn.Module):
def __init__(
self,
dim: int,
num_heads: int = 8,
qkv_bias: bool = True,
proj_bias: bool = True,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
norm_layer: nn.Module = nn.LayerNorm,
qk_norm: bool = False,
fused_attn: bool = True, # use F.scaled_dot_product_attention or not
rope=None,
) -> None:
super().__init__()
assert dim % num_heads == 0, "dim should be divisible by num_heads"
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.scale = self.head_dim**-0.5
self.fused_attn = fused_attn
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim, bias=proj_bias)
self.proj_drop = nn.Dropout(proj_drop)
self.rope = rope
def forward(self, x: Tensor, pos=None) -> Tensor:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
q, k, v = qkv.unbind(0)
q, k = self.q_norm(q), self.k_norm(k)
if self.rope is not None:
q = self.rope(q, pos)
k = self.rope(k, pos)
if self.fused_attn:
x = F.scaled_dot_product_attention(
q,
k,
v,
dropout_p=self.attn_drop.p if self.training else 0.0,
)
else:
q = q * self.scale
attn = q @ k.transpose(-2, -1)
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = attn @ v
x = x.transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class MemEffAttention(Attention):
def forward(self, x: Tensor, attn_bias=None, pos=None) -> Tensor:
assert pos is None
if not XFORMERS_AVAILABLE:
if attn_bias is not None:
raise AssertionError("xFormers is required for using nested tensors")
return super().forward(x)
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
q, k, v = unbind(qkv, 2)
x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
x = x.reshape([B, N, C])
x = self.proj(x)
x = self.proj_drop(x)
return x
vggt/vggt/layers/block.py
+259
View File
@@ -0,0 +1,259 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/patch_embed.py
import logging
import os
from typing import Callable, List, Any, Tuple, Dict
import warnings
import torch
from torch import nn, Tensor
from .attention import Attention
from .drop_path import DropPath
from .layer_scale import LayerScale
from .mlp import Mlp
XFORMERS_AVAILABLE = False
class Block(nn.Module):
def __init__(
self,
dim: int,
num_heads: int,
mlp_ratio: float = 4.0,
qkv_bias: bool = True,
proj_bias: bool = True,
ffn_bias: bool = True,
drop: float = 0.0,
attn_drop: float = 0.0,
init_values=None,
drop_path: float = 0.0,
act_layer: Callable[..., nn.Module] = nn.GELU,
norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
attn_class: Callable[..., nn.Module] = Attention,
ffn_layer: Callable[..., nn.Module] = Mlp,
qk_norm: bool = False,
fused_attn: bool = True, # use F.scaled_dot_product_attention or not
rope=None,
) -> None:
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = attn_class(
dim,
num_heads=num_heads,
qkv_bias=qkv_bias,
proj_bias=proj_bias,
attn_drop=attn_drop,
proj_drop=drop,
qk_norm=qk_norm,
fused_attn=fused_attn,
rope=rope,
)
self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
self.drop_path1 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = ffn_layer(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
bias=ffn_bias,
)
self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
self.drop_path2 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.sample_drop_ratio = drop_path
def forward(self, x: Tensor, pos=None) -> Tensor:
def attn_residual_func(x: Tensor, pos=None) -> Tensor:
return self.ls1(self.attn(self.norm1(x), pos=pos))
def ffn_residual_func(x: Tensor) -> Tensor:
return self.ls2(self.mlp(self.norm2(x)))
if self.training and self.sample_drop_ratio > 0.1:
# the overhead is compensated only for a drop path rate larger than 0.1
x = drop_add_residual_stochastic_depth(
x,
pos=pos,
residual_func=attn_residual_func,
sample_drop_ratio=self.sample_drop_ratio,
)
x = drop_add_residual_stochastic_depth(
x,
residual_func=ffn_residual_func,
sample_drop_ratio=self.sample_drop_ratio,
)
elif self.training and self.sample_drop_ratio > 0.0:
x = x + self.drop_path1(attn_residual_func(x, pos=pos))
x = x + self.drop_path1(ffn_residual_func(x)) # FIXME: drop_path2
else:
x = x + attn_residual_func(x, pos=pos)
x = x + ffn_residual_func(x)
return x
def drop_add_residual_stochastic_depth(
x: Tensor,
residual_func: Callable[[Tensor], Tensor],
sample_drop_ratio: float = 0.0,
pos=None,
) -> Tensor:
# 1) extract subset using permutation
b, n, d = x.shape
sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
x_subset = x[brange]
# 2) apply residual_func to get residual
if pos is not None:
# if necessary, apply rope to the subset
pos = pos[brange]
residual = residual_func(x_subset, pos=pos)
else:
residual = residual_func(x_subset)
x_flat = x.flatten(1)
residual = residual.flatten(1)
residual_scale_factor = b / sample_subset_size
# 3) add the residual
x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
return x_plus_residual.view_as(x)
def get_branges_scales(x, sample_drop_ratio=0.0):
b, n, d = x.shape
sample_subset_size = max(int(b * (1 - sample_drop_ratio)), 1)
brange = (torch.randperm(b, device=x.device))[:sample_subset_size]
residual_scale_factor = b / sample_subset_size
return brange, residual_scale_factor
def add_residual(x, brange, residual, residual_scale_factor, scaling_vector=None):
if scaling_vector is None:
x_flat = x.flatten(1)
residual = residual.flatten(1)
x_plus_residual = torch.index_add(x_flat, 0, brange, residual.to(dtype=x.dtype), alpha=residual_scale_factor)
else:
x_plus_residual = scaled_index_add(
x, brange, residual.to(dtype=x.dtype), scaling=scaling_vector, alpha=residual_scale_factor
)
return x_plus_residual
attn_bias_cache: Dict[Tuple, Any] = {}
def get_attn_bias_and_cat(x_list, branges=None):
"""
this will perform the index select, cat the tensors, and provide the attn_bias from cache
"""
batch_sizes = [b.shape[0] for b in branges] if branges is not None else [x.shape[0] for x in x_list]
all_shapes = tuple((b, x.shape[1]) for b, x in zip(batch_sizes, x_list))
if all_shapes not in attn_bias_cache.keys():
seqlens = []
for b, x in zip(batch_sizes, x_list):
for _ in range(b):
seqlens.append(x.shape[1])
attn_bias = fmha.BlockDiagonalMask.from_seqlens(seqlens)
attn_bias._batch_sizes = batch_sizes
attn_bias_cache[all_shapes] = attn_bias
if branges is not None:
cat_tensors = index_select_cat([x.flatten(1) for x in x_list], branges).view(1, -1, x_list[0].shape[-1])
else:
tensors_bs1 = tuple(x.reshape([1, -1, *x.shape[2:]]) for x in x_list)
cat_tensors = torch.cat(tensors_bs1, dim=1)
return attn_bias_cache[all_shapes], cat_tensors
def drop_add_residual_stochastic_depth_list(
x_list: List[Tensor],
residual_func: Callable[[Tensor, Any], Tensor],
sample_drop_ratio: float = 0.0,
scaling_vector=None,
) -> Tensor:
# 1) generate random set of indices for dropping samples in the batch
branges_scales = [get_branges_scales(x, sample_drop_ratio=sample_drop_ratio) for x in x_list]
branges = [s[0] for s in branges_scales]
residual_scale_factors = [s[1] for s in branges_scales]
# 2) get attention bias and index+concat the tensors
attn_bias, x_cat = get_attn_bias_and_cat(x_list, branges)
# 3) apply residual_func to get residual, and split the result
residual_list = attn_bias.split(residual_func(x_cat, attn_bias=attn_bias)) # type: ignore
outputs = []
for x, brange, residual, residual_scale_factor in zip(x_list, branges, residual_list, residual_scale_factors):
outputs.append(add_residual(x, brange, residual, residual_scale_factor, scaling_vector).view_as(x))
return outputs
class NestedTensorBlock(Block):
def forward_nested(self, x_list: List[Tensor]) -> List[Tensor]:
"""
x_list contains a list of tensors to nest together and run
"""
assert isinstance(self.attn, MemEffAttention)
if self.training and self.sample_drop_ratio > 0.0:
def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
return self.attn(self.norm1(x), attn_bias=attn_bias)
def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
return self.mlp(self.norm2(x))
x_list = drop_add_residual_stochastic_depth_list(
x_list,
residual_func=attn_residual_func,
sample_drop_ratio=self.sample_drop_ratio,
scaling_vector=self.ls1.gamma if isinstance(self.ls1, LayerScale) else None,
)
x_list = drop_add_residual_stochastic_depth_list(
x_list,
residual_func=ffn_residual_func,
sample_drop_ratio=self.sample_drop_ratio,
scaling_vector=self.ls2.gamma if isinstance(self.ls1, LayerScale) else None,
)
return x_list
else:
def attn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
return self.ls1(self.attn(self.norm1(x), attn_bias=attn_bias))
def ffn_residual_func(x: Tensor, attn_bias=None) -> Tensor:
return self.ls2(self.mlp(self.norm2(x)))
attn_bias, x = get_attn_bias_and_cat(x_list)
x = x + attn_residual_func(x, attn_bias=attn_bias)
x = x + ffn_residual_func(x)
return attn_bias.split(x)
def forward(self, x_or_x_list):
if isinstance(x_or_x_list, Tensor):
return super().forward(x_or_x_list)
elif isinstance(x_or_x_list, list):
if not XFORMERS_AVAILABLE:
raise AssertionError("xFormers is required for using nested tensors")
return self.forward_nested(x_or_x_list)
else:
raise AssertionError
vggt/vggt/layers/drop_path.py
+34
View File
@@ -0,0 +1,34 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/drop.py
from torch import nn
def drop_path(x, drop_prob: float = 0.0, training: bool = False):
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets
random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
if keep_prob > 0.0:
random_tensor.div_(keep_prob)
output = x * random_tensor
return output
class DropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training)
vggt/vggt/layers/layer_scale.py
+27
View File
@@ -0,0 +1,27 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# Modified from: https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py#L103-L110
from typing import Union
import torch
from torch import Tensor
from torch import nn
class LayerScale(nn.Module):
def __init__(
self,
dim: int,
init_values: Union[float, Tensor] = 1e-5,
inplace: bool = False,
) -> None:
super().__init__()
self.inplace = inplace
self.gamma = nn.Parameter(init_values * torch.ones(dim))
def forward(self, x: Tensor) -> Tensor:
return x.mul_(self.gamma) if self.inplace else x * self.gamma
vggt/vggt/layers/mlp.py
+40
View File
@@ -0,0 +1,40 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/mlp.py
from typing import Callable, Optional
from torch import Tensor, nn
class Mlp(nn.Module):
def __init__(
self,
in_features: int,
hidden_features: Optional[int] = None,
out_features: Optional[int] = None,
act_layer: Callable[..., nn.Module] = nn.GELU,
drop: float = 0.0,
bias: bool = True,
) -> None:
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias)
self.drop = nn.Dropout(drop)
def forward(self, x: Tensor) -> Tensor:
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
vggt/vggt/layers/patch_embed.py
+88
View File
@@ -0,0 +1,88 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/layers/patch_embed.py
from typing import Callable, Optional, Tuple, Union
from torch import Tensor
import torch.nn as nn
def make_2tuple(x):
if isinstance(x, tuple):
assert len(x) == 2
return x
assert isinstance(x, int)
return (x, x)
class PatchEmbed(nn.Module):
"""
2D image to patch embedding: (B,C,H,W) -> (B,N,D)
Args:
img_size: Image size.
patch_size: Patch token size.
in_chans: Number of input image channels.
embed_dim: Number of linear projection output channels.
norm_layer: Normalization layer.
"""
def __init__(
self,
img_size: Union[int, Tuple[int, int]] = 224,
patch_size: Union[int, Tuple[int, int]] = 16,
in_chans: int = 3,
embed_dim: int = 768,
norm_layer: Optional[Callable] = None,
flatten_embedding: bool = True,
) -> None:
super().__init__()
image_HW = make_2tuple(img_size)
patch_HW = make_2tuple(patch_size)
patch_grid_size = (
image_HW[0] // patch_HW[0],
image_HW[1] // patch_HW[1],
)
self.img_size = image_HW
self.patch_size = patch_HW
self.patches_resolution = patch_grid_size
self.num_patches = patch_grid_size[0] * patch_grid_size[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
self.flatten_embedding = flatten_embedding
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_HW, stride=patch_HW)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x: Tensor) -> Tensor:
_, _, H, W = x.shape
patch_H, patch_W = self.patch_size
assert H % patch_H == 0, f"Input image height {H} is not a multiple of patch height {patch_H}"
assert W % patch_W == 0, f"Input image width {W} is not a multiple of patch width: {patch_W}"
x = self.proj(x) # B C H W
H, W = x.size(2), x.size(3)
x = x.flatten(2).transpose(1, 2) # B HW C
x = self.norm(x)
if not self.flatten_embedding:
x = x.reshape(-1, H, W, self.embed_dim) # B H W C
return x
def flops(self) -> float:
Ho, Wo = self.patches_resolution
flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
if self.norm is not None:
flops += Ho * Wo * self.embed_dim
return flops
vggt/vggt/layers/rope.py
+188
View File
@@ -0,0 +1,188 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# Implementation of 2D Rotary Position Embeddings (RoPE).
# This module provides a clean implementation of 2D Rotary Position Embeddings,
# which extends the original RoPE concept to handle 2D spatial positions.
# Inspired by:
# https://github.com/meta-llama/codellama/blob/main/llama/model.py
# https://github.com/naver-ai/rope-vit
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, Tuple
class PositionGetter:
"""Generates and caches 2D spatial positions for patches in a grid.
This class efficiently manages the generation of spatial coordinates for patches
in a 2D grid, caching results to avoid redundant computations.
Attributes:
position_cache: Dictionary storing precomputed position tensors for different
grid dimensions.
"""
def __init__(self):
"""Initializes the position generator with an empty cache."""
self.position_cache: Dict[Tuple[int, int], torch.Tensor] = {}
def __call__(self, batch_size: int, height: int, width: int, device: torch.device) -> torch.Tensor:
"""Generates spatial positions for a batch of patches.
Args:
batch_size: Number of samples in the batch.
height: Height of the grid in patches.
width: Width of the grid in patches.
device: Target device for the position tensor.
Returns:
Tensor of shape (batch_size, height*width, 2) containing y,x coordinates
for each position in the grid, repeated for each batch item.
"""
if (height, width) not in self.position_cache:
y_coords = torch.arange(height, device=device)
x_coords = torch.arange(width, device=device)
positions = torch.cartesian_prod(y_coords, x_coords)
self.position_cache[height, width] = positions
cached_positions = self.position_cache[height, width]
return cached_positions.view(1, height * width, 2).expand(batch_size, -1, -1).clone()
class RotaryPositionEmbedding2D(nn.Module):
"""2D Rotary Position Embedding implementation.
This module applies rotary position embeddings to input tokens based on their
2D spatial positions. It handles the position-dependent rotation of features
separately for vertical and horizontal dimensions.
Args:
frequency: Base frequency for the position embeddings. Default: 100.0
scaling_factor: Scaling factor for frequency computation. Default: 1.0
Attributes:
base_frequency: Base frequency for computing position embeddings.
scaling_factor: Factor to scale the computed frequencies.
frequency_cache: Cache for storing precomputed frequency components.
"""
def __init__(self, frequency: float = 100.0, scaling_factor: float = 1.0):
"""Initializes the 2D RoPE module."""
super().__init__()
self.base_frequency = frequency
self.scaling_factor = scaling_factor
self.frequency_cache: Dict[Tuple, Tuple[torch.Tensor, torch.Tensor]] = {}
def _compute_frequency_components(
self, dim: int, seq_len: int, device: torch.device, dtype: torch.dtype
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Computes frequency components for rotary embeddings.
Args:
dim: Feature dimension (must be even).
seq_len: Maximum sequence length.
device: Target device for computations.
dtype: Data type for the computed tensors.
Returns:
Tuple of (cosine, sine) tensors for frequency components.
"""
cache_key = (dim, seq_len, device, dtype)
if cache_key not in self.frequency_cache:
# Compute frequency bands
exponents = torch.arange(0, dim, 2, device=device).float() / dim
inv_freq = 1.0 / (self.base_frequency**exponents)
# Generate position-dependent frequencies
positions = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
angles = torch.einsum("i,j->ij", positions, inv_freq)
# Compute and cache frequency components
angles = angles.to(dtype)
angles = torch.cat((angles, angles), dim=-1)
cos_components = angles.cos().to(dtype)
sin_components = angles.sin().to(dtype)
self.frequency_cache[cache_key] = (cos_components, sin_components)
return self.frequency_cache[cache_key]
@staticmethod
def _rotate_features(x: torch.Tensor) -> torch.Tensor:
"""Performs feature rotation by splitting and recombining feature dimensions.
Args:
x: Input tensor to rotate.
Returns:
Rotated feature tensor.
"""
feature_dim = x.shape[-1]
x1, x2 = x[..., : feature_dim // 2], x[..., feature_dim // 2 :]
return torch.cat((-x2, x1), dim=-1)
def _apply_1d_rope(
self, tokens: torch.Tensor, positions: torch.Tensor, cos_comp: torch.Tensor, sin_comp: torch.Tensor
) -> torch.Tensor:
"""Applies 1D rotary position embeddings along one dimension.
Args:
tokens: Input token features.
positions: Position indices.
cos_comp: Cosine components for rotation.
sin_comp: Sine components for rotation.
Returns:
Tokens with applied rotary position embeddings.
"""
# Embed positions with frequency components
cos = F.embedding(positions, cos_comp)[:, None, :, :]
sin = F.embedding(positions, sin_comp)[:, None, :, :]
# Apply rotation
return (tokens * cos) + (self._rotate_features(tokens) * sin)
def forward(self, tokens: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
"""Applies 2D rotary position embeddings to input tokens.
Args:
tokens: Input tensor of shape (batch_size, n_heads, n_tokens, dim).
The feature dimension (dim) must be divisible by 4.
positions: Position tensor of shape (batch_size, n_tokens, 2) containing
the y and x coordinates for each token.
Returns:
Tensor of same shape as input with applied 2D rotary position embeddings.
Raises:
AssertionError: If input dimensions are invalid or positions are malformed.
"""
# Validate inputs
assert tokens.size(-1) % 2 == 0, "Feature dimension must be even"
assert positions.ndim == 3 and positions.shape[-1] == 2, "Positions must have shape (batch_size, n_tokens, 2)"
# Compute feature dimension for each spatial direction
feature_dim = tokens.size(-1) // 2
# Get frequency components
max_position = int(positions.max()) + 1
cos_comp, sin_comp = self._compute_frequency_components(feature_dim, max_position, tokens.device, tokens.dtype)
# Split features for vertical and horizontal processing
vertical_features, horizontal_features = tokens.chunk(2, dim=-1)
# Apply RoPE separately for each dimension
vertical_features = self._apply_1d_rope(vertical_features, positions[..., 0], cos_comp, sin_comp)
horizontal_features = self._apply_1d_rope(horizontal_features, positions[..., 1], cos_comp, sin_comp)
# Combine processed features
return torch.cat((vertical_features, horizontal_features), dim=-1)
vggt/vggt/layers/swiglu_ffn.py
+72
View File
@@ -0,0 +1,72 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
import os
from typing import Callable, Optional
import warnings
from torch import Tensor, nn
import torch.nn.functional as F
class SwiGLUFFN(nn.Module):
def __init__(
self,
in_features: int,
hidden_features: Optional[int] = None,
out_features: Optional[int] = None,
act_layer: Callable[..., nn.Module] = None,
drop: float = 0.0,
bias: bool = True,
) -> None:
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.w12 = nn.Linear(in_features, 2 * hidden_features, bias=bias)
self.w3 = nn.Linear(hidden_features, out_features, bias=bias)
def forward(self, x: Tensor) -> Tensor:
x12 = self.w12(x)
x1, x2 = x12.chunk(2, dim=-1)
hidden = F.silu(x1) * x2
return self.w3(hidden)
XFORMERS_ENABLED = os.environ.get("XFORMERS_DISABLED") is None
# try:
# if XFORMERS_ENABLED:
# from xformers.ops import SwiGLU
# XFORMERS_AVAILABLE = True
# warnings.warn("xFormers is available (SwiGLU)")
# else:
# warnings.warn("xFormers is disabled (SwiGLU)")
# raise ImportError
# except ImportError:
SwiGLU = SwiGLUFFN
XFORMERS_AVAILABLE = False
# warnings.warn("xFormers is not available (SwiGLU)")
class SwiGLUFFNFused(SwiGLU):
def __init__(
self,
in_features: int,
hidden_features: Optional[int] = None,
out_features: Optional[int] = None,
act_layer: Callable[..., nn.Module] = None,
drop: float = 0.0,
bias: bool = True,
) -> None:
out_features = out_features or in_features
hidden_features = hidden_features or in_features
hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
super().__init__(
in_features=in_features,
hidden_features=hidden_features,
out_features=out_features,
bias=bias,
)
vggt/vggt/layers/vision_transformer.py
+407
View File
@@ -0,0 +1,407 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/main/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py
from functools import partial
import math
import logging
from typing import Sequence, Tuple, Union, Callable
import torch
import torch.nn as nn
from torch.utils.checkpoint import checkpoint
from torch.nn.init import trunc_normal_
from . import Mlp, PatchEmbed, SwiGLUFFNFused, MemEffAttention, NestedTensorBlock as Block
logger = logging.getLogger("dinov2")
def named_apply(fn: Callable, module: nn.Module, name="", depth_first=True, include_root=False) -> nn.Module:
if not depth_first and include_root:
fn(module=module, name=name)
for child_name, child_module in module.named_children():
child_name = ".".join((name, child_name)) if name else child_name
named_apply(fn=fn, module=child_module, name=child_name, depth_first=depth_first, include_root=True)
if depth_first and include_root:
fn(module=module, name=name)
return module
class BlockChunk(nn.ModuleList):
def forward(self, x):
for b in self:
x = b(x)
return x
class DinoVisionTransformer(nn.Module):
def __init__(
self,
img_size=224,
patch_size=16,
in_chans=3,
embed_dim=768,
depth=12,
num_heads=12,
mlp_ratio=4.0,
qkv_bias=True,
ffn_bias=True,
proj_bias=True,
drop_path_rate=0.0,
drop_path_uniform=False,
init_values=None, # for layerscale: None or 0 => no layerscale
embed_layer=PatchEmbed,
act_layer=nn.GELU,
block_fn=Block,
ffn_layer="mlp",
block_chunks=1,
num_register_tokens=0,
interpolate_antialias=False,
interpolate_offset=0.1,
qk_norm=False,
):
"""
Args:
img_size (int, tuple): input image size
patch_size (int, tuple): patch size
in_chans (int): number of input channels
embed_dim (int): embedding dimension
depth (int): depth of transformer
num_heads (int): number of attention heads
mlp_ratio (int): ratio of mlp hidden dim to embedding dim
qkv_bias (bool): enable bias for qkv if True
proj_bias (bool): enable bias for proj in attn if True
ffn_bias (bool): enable bias for ffn if True
drop_path_rate (float): stochastic depth rate
drop_path_uniform (bool): apply uniform drop rate across blocks
weight_init (str): weight init scheme
init_values (float): layer-scale init values
embed_layer (nn.Module): patch embedding layer
act_layer (nn.Module): MLP activation layer
block_fn (nn.Module): transformer block class
ffn_layer (str): "mlp", "swiglu", "swiglufused" or "identity"
block_chunks: (int) split block sequence into block_chunks units for FSDP wrap
num_register_tokens: (int) number of extra cls tokens (so-called "registers")
interpolate_antialias: (str) flag to apply anti-aliasing when interpolating positional embeddings
interpolate_offset: (float) work-around offset to apply when interpolating positional embeddings
"""
super().__init__()
norm_layer = partial(nn.LayerNorm, eps=1e-6)
# tricky but makes it work
self.use_checkpoint = False
#
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
self.num_tokens = 1
self.n_blocks = depth
self.num_heads = num_heads
self.patch_size = patch_size
self.num_register_tokens = num_register_tokens
self.interpolate_antialias = interpolate_antialias
self.interpolate_offset = interpolate_offset
self.patch_embed = embed_layer(img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
num_patches = self.patch_embed.num_patches
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
assert num_register_tokens >= 0
self.register_tokens = (
nn.Parameter(torch.zeros(1, num_register_tokens, embed_dim)) if num_register_tokens else None
)
if drop_path_uniform is True:
dpr = [drop_path_rate] * depth
else:
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
if ffn_layer == "mlp":
logger.info("using MLP layer as FFN")
ffn_layer = Mlp
elif ffn_layer == "swiglufused" or ffn_layer == "swiglu":
logger.info("using SwiGLU layer as FFN")
ffn_layer = SwiGLUFFNFused
elif ffn_layer == "identity":
logger.info("using Identity layer as FFN")
def f(*args, **kwargs):
return nn.Identity()
ffn_layer = f
else:
raise NotImplementedError
blocks_list = [
block_fn(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
proj_bias=proj_bias,
ffn_bias=ffn_bias,
drop_path=dpr[i],
norm_layer=norm_layer,
act_layer=act_layer,
ffn_layer=ffn_layer,
init_values=init_values,
qk_norm=qk_norm,
)
for i in range(depth)
]
if block_chunks > 0:
self.chunked_blocks = True
chunked_blocks = []
chunksize = depth // block_chunks
for i in range(0, depth, chunksize):
# this is to keep the block index consistent if we chunk the block list
chunked_blocks.append([nn.Identity()] * i + blocks_list[i : i + chunksize])
self.blocks = nn.ModuleList([BlockChunk(p) for p in chunked_blocks])
else:
self.chunked_blocks = False
self.blocks = nn.ModuleList(blocks_list)
self.norm = norm_layer(embed_dim)
self.head = nn.Identity()
self.mask_token = nn.Parameter(torch.zeros(1, embed_dim))
self.init_weights()
def init_weights(self):
trunc_normal_(self.pos_embed, std=0.02)
nn.init.normal_(self.cls_token, std=1e-6)
if self.register_tokens is not None:
nn.init.normal_(self.register_tokens, std=1e-6)
named_apply(init_weights_vit_timm, self)
def interpolate_pos_encoding(self, x, w, h):
previous_dtype = x.dtype
npatch = x.shape[1] - 1
N = self.pos_embed.shape[1] - 1
if npatch == N and w == h:
return self.pos_embed
pos_embed = self.pos_embed.float()
class_pos_embed = pos_embed[:, 0]
patch_pos_embed = pos_embed[:, 1:]
dim = x.shape[-1]
w0 = w // self.patch_size
h0 = h // self.patch_size
M = int(math.sqrt(N)) # Recover the number of patches in each dimension
assert N == M * M
kwargs = {}
if self.interpolate_offset:
# Historical kludge: add a small number to avoid floating point error in the interpolation, see https://github.com/facebookresearch/dino/issues/8
# Note: still needed for backward-compatibility, the underlying operators are using both output size and scale factors
sx = float(w0 + self.interpolate_offset) / M
sy = float(h0 + self.interpolate_offset) / M
kwargs["scale_factor"] = (sx, sy)
else:
# Simply specify an output size instead of a scale factor
kwargs["size"] = (w0, h0)
patch_pos_embed = nn.functional.interpolate(
patch_pos_embed.reshape(1, M, M, dim).permute(0, 3, 1, 2),
mode="bicubic",
antialias=self.interpolate_antialias,
**kwargs,
)
assert (w0, h0) == patch_pos_embed.shape[-2:]
patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1).to(previous_dtype)
def prepare_tokens_with_masks(self, x, masks=None):
B, nc, w, h = x.shape
x = self.patch_embed(x)
if masks is not None:
x = torch.where(masks.unsqueeze(-1), self.mask_token.to(x.dtype).unsqueeze(0), x)
x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
x = x + self.interpolate_pos_encoding(x, w, h)
if self.register_tokens is not None:
x = torch.cat(
(
x[:, :1],
self.register_tokens.expand(x.shape[0], -1, -1),
x[:, 1:],
),
dim=1,
)
return x
def forward_features_list(self, x_list, masks_list):
x = [self.prepare_tokens_with_masks(x, masks) for x, masks in zip(x_list, masks_list)]
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint(blk, x, use_reentrant=self.use_reentrant)
else:
x = blk(x)
all_x = x
output = []
for x, masks in zip(all_x, masks_list):
x_norm = self.norm(x)
output.append(
{
"x_norm_clstoken": x_norm[:, 0],
"x_norm_regtokens": x_norm[:, 1 : self.num_register_tokens + 1],
"x_norm_patchtokens": x_norm[:, self.num_register_tokens + 1 :],
"x_prenorm": x,
"masks": masks,
}
)
return output
def forward_features(self, x, masks=None):
if isinstance(x, list):
return self.forward_features_list(x, masks)
x = self.prepare_tokens_with_masks(x, masks)
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint(blk, x, use_reentrant=self.use_reentrant)
else:
x = blk(x)
x_norm = self.norm(x)
return {
"x_norm_clstoken": x_norm[:, 0],
"x_norm_regtokens": x_norm[:, 1 : self.num_register_tokens + 1],
"x_norm_patchtokens": x_norm[:, self.num_register_tokens + 1 :],
"x_prenorm": x,
"masks": masks,
}
def _get_intermediate_layers_not_chunked(self, x, n=1):
x = self.prepare_tokens_with_masks(x)
# If n is an int, take the n last blocks. If it's a list, take them
output, total_block_len = [], len(self.blocks)
blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
for i, blk in enumerate(self.blocks):
x = blk(x)
if i in blocks_to_take:
output.append(x)
assert len(output) == len(blocks_to_take), f"only {len(output)} / {len(blocks_to_take)} blocks found"
return output
def _get_intermediate_layers_chunked(self, x, n=1):
x = self.prepare_tokens_with_masks(x)
output, i, total_block_len = [], 0, len(self.blocks[-1])
# If n is an int, take the n last blocks. If it's a list, take them
blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
for block_chunk in self.blocks:
for blk in block_chunk[i:]: # Passing the nn.Identity()
x = blk(x)
if i in blocks_to_take:
output.append(x)
i += 1
assert len(output) == len(blocks_to_take), f"only {len(output)} / {len(blocks_to_take)} blocks found"
return output
def get_intermediate_layers(
self,
x: torch.Tensor,
n: Union[int, Sequence] = 1, # Layers or n last layers to take
reshape: bool = False,
return_class_token: bool = False,
norm=True,
) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]]]:
if self.chunked_blocks:
outputs = self._get_intermediate_layers_chunked(x, n)
else:
outputs = self._get_intermediate_layers_not_chunked(x, n)
if norm:
outputs = [self.norm(out) for out in outputs]
class_tokens = [out[:, 0] for out in outputs]
outputs = [out[:, 1 + self.num_register_tokens :] for out in outputs]
if reshape:
B, _, w, h = x.shape
outputs = [
out.reshape(B, w // self.patch_size, h // self.patch_size, -1).permute(0, 3, 1, 2).contiguous()
for out in outputs
]
if return_class_token:
return tuple(zip(outputs, class_tokens))
return tuple(outputs)
def forward(self, *args, is_training=True, **kwargs):
ret = self.forward_features(*args, **kwargs)
if is_training:
return ret
else:
return self.head(ret["x_norm_clstoken"])
def init_weights_vit_timm(module: nn.Module, name: str = ""):
"""ViT weight initialization, original timm impl (for reproducibility)"""
if isinstance(module, nn.Linear):
trunc_normal_(module.weight, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
def vit_small(patch_size=16, num_register_tokens=0, **kwargs):
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=384,
depth=12,
num_heads=6,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
def vit_base(patch_size=16, num_register_tokens=0, **kwargs):
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=768,
depth=12,
num_heads=12,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
def vit_large(patch_size=16, num_register_tokens=0, **kwargs):
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=1024,
depth=24,
num_heads=16,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
def vit_giant2(patch_size=16, num_register_tokens=0, **kwargs):
"""
Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64
"""
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=1536,
depth=40,
num_heads=24,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
vggt/vggt/models/aggregator.py
+331
View File
@@ -0,0 +1,331 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple, Union, List, Dict, Any
from ..layers import PatchEmbed
from ..layers.block import Block
from ..layers.rope import RotaryPositionEmbedding2D, PositionGetter
from ..layers.vision_transformer import vit_small, vit_base, vit_large, vit_giant2
logger = logging.getLogger(__name__)
_RESNET_MEAN = [0.485, 0.456, 0.406]
_RESNET_STD = [0.229, 0.224, 0.225]
class Aggregator(nn.Module):
"""
The Aggregator applies alternating-attention over input frames,
as described in VGGT: Visual Geometry Grounded Transformer.
Args:
img_size (int): Image size in pixels.
patch_size (int): Size of each patch for PatchEmbed.
embed_dim (int): Dimension of the token embeddings.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
mlp_ratio (float): Ratio of MLP hidden dim to embedding dim.
num_register_tokens (int): Number of register tokens.
block_fn (nn.Module): The block type used for attention (Block by default).
qkv_bias (bool): Whether to include bias in QKV projections.
proj_bias (bool): Whether to include bias in the output projection.
ffn_bias (bool): Whether to include bias in MLP layers.
patch_embed (str): Type of patch embed. e.g., "conv" or "dinov2_vitl14_reg".
aa_order (list[str]): The order of alternating attention, e.g. ["frame", "global"].
aa_block_size (int): How many blocks to group under each attention type before switching. If not necessary, set to 1.
qk_norm (bool): Whether to apply QK normalization.
rope_freq (int): Base frequency for rotary embedding. -1 to disable.
init_values (float): Init scale for layer scale.
"""
def __init__(
self,
img_size=518,
patch_size=14,
embed_dim=1024,
depth=24,
num_heads=16,
mlp_ratio=4.0,
num_register_tokens=4,
block_fn=Block,
qkv_bias=True,
proj_bias=True,
ffn_bias=True,
patch_embed="dinov2_vitl14_reg",
aa_order=["frame", "global"],
aa_block_size=1,
qk_norm=True,
rope_freq=100,
init_values=0.01,
):
super().__init__()
self.__build_patch_embed__(patch_embed, img_size, patch_size, num_register_tokens, embed_dim=embed_dim)
# Initialize rotary position embedding if frequency > 0
self.rope = RotaryPositionEmbedding2D(frequency=rope_freq) if rope_freq > 0 else None
self.position_getter = PositionGetter() if self.rope is not None else None
self.frame_blocks = nn.ModuleList(
[
block_fn(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
proj_bias=proj_bias,
ffn_bias=ffn_bias,
init_values=init_values,
qk_norm=qk_norm,
rope=self.rope,
)
for _ in range(depth)
]
)
self.global_blocks = nn.ModuleList(
[
block_fn(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
proj_bias=proj_bias,
ffn_bias=ffn_bias,
init_values=init_values,
qk_norm=qk_norm,
rope=self.rope,
)
for _ in range(depth)
]
)
self.depth = depth
self.aa_order = aa_order
self.patch_size = patch_size
self.aa_block_size = aa_block_size
# Validate that depth is divisible by aa_block_size
if self.depth % self.aa_block_size != 0:
raise ValueError(f"depth ({depth}) must be divisible by aa_block_size ({aa_block_size})")
self.aa_block_num = self.depth // self.aa_block_size
# Note: We have two camera tokens, one for the first frame and one for the rest
# The same applies for register tokens
self.camera_token = nn.Parameter(torch.randn(1, 2, 1, embed_dim))
self.register_token = nn.Parameter(torch.randn(1, 2, num_register_tokens, embed_dim))
# The patch tokens start after the camera and register tokens
self.patch_start_idx = 1 + num_register_tokens
# Initialize parameters with small values
nn.init.normal_(self.camera_token, std=1e-6)
nn.init.normal_(self.register_token, std=1e-6)
# Register normalization constants as buffers
for name, value in (
("_resnet_mean", _RESNET_MEAN),
("_resnet_std", _RESNET_STD),
):
self.register_buffer(
name,
torch.FloatTensor(value).view(1, 1, 3, 1, 1),
persistent=False,
)
def __build_patch_embed__(
self,
patch_embed,
img_size,
patch_size,
num_register_tokens,
interpolate_antialias=True,
interpolate_offset=0.0,
block_chunks=0,
init_values=1.0,
embed_dim=1024,
):
"""
Build the patch embed layer. If 'conv', we use a
simple PatchEmbed conv layer. Otherwise, we use a vision transformer.
"""
if "conv" in patch_embed:
self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=3, embed_dim=embed_dim)
else:
vit_models = {
"dinov2_vitl14_reg": vit_large,
"dinov2_vitb14_reg": vit_base,
"dinov2_vits14_reg": vit_small,
"dinov2_vitg2_reg": vit_giant2,
}
self.patch_embed = vit_models[patch_embed](
img_size=img_size,
patch_size=patch_size,
num_register_tokens=num_register_tokens,
interpolate_antialias=interpolate_antialias,
interpolate_offset=interpolate_offset,
block_chunks=block_chunks,
init_values=init_values,
)
# Disable gradient updates for mask token
if hasattr(self.patch_embed, "mask_token"):
self.patch_embed.mask_token.requires_grad_(False)
def forward(
self,
images: torch.Tensor,
) -> Tuple[List[torch.Tensor], int]:
"""
Args:
images (torch.Tensor): Input images with shape [B, S, 3, H, W], in range [0, 1].
B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
Returns:
(list[torch.Tensor], int):
The list of outputs from the attention blocks,
and the patch_start_idx indicating where patch tokens begin.
"""
B, S, C_in, H, W = images.shape
if C_in != 3:
raise ValueError(f"Expected 3 input channels, got {C_in}")
# Normalize images and reshape for patch embed
images = (images - self._resnet_mean) / self._resnet_std
# Reshape to [B*S, C, H, W] for patch embedding
images = images.view(B * S, C_in, H, W)
patch_tokens = self.patch_embed(images)
if isinstance(patch_tokens, dict):
patch_tokens = patch_tokens["x_norm_patchtokens"]
_, P, C = patch_tokens.shape
# Expand camera and register tokens to match batch size and sequence length
camera_token = slice_expand_and_flatten(self.camera_token, B, S)
register_token = slice_expand_and_flatten(self.register_token, B, S)
# Concatenate special tokens with patch tokens
tokens = torch.cat([camera_token, register_token, patch_tokens], dim=1)
pos = None
if self.rope is not None:
pos = self.position_getter(B * S, H // self.patch_size, W // self.patch_size, device=images.device)
if self.patch_start_idx > 0:
# do not use position embedding for special tokens (camera and register tokens)
# so set pos to 0 for the special tokens
pos = pos + 1
pos_special = torch.zeros(B * S, self.patch_start_idx, 2).to(images.device).to(pos.dtype)
pos = torch.cat([pos_special, pos], dim=1)
# update P because we added special tokens
_, P, C = tokens.shape
frame_idx = 0
global_idx = 0
output_list = []
for _ in range(self.aa_block_num):
for attn_type in self.aa_order:
if attn_type == "frame":
tokens, frame_idx, frame_intermediates = self._process_frame_attention(
tokens, B, S, P, C, frame_idx, pos=pos
)
elif attn_type == "global":
tokens, global_idx, global_intermediates = self._process_global_attention(
tokens, B, S, P, C, global_idx, pos=pos
)
else:
raise ValueError(f"Unknown attention type: {attn_type}")
for i in range(len(frame_intermediates)):
# concat frame and global intermediates, [B x S x P x 2C]
concat_inter = torch.cat([frame_intermediates[i], global_intermediates[i]], dim=-1)
output_list.append(concat_inter)
del concat_inter
del frame_intermediates
del global_intermediates
return output_list, self.patch_start_idx
def _process_frame_attention(self, tokens, B, S, P, C, frame_idx, pos=None):
"""
Process frame attention blocks. We keep tokens in shape (B*S, P, C).
"""
# If needed, reshape tokens or positions:
if tokens.shape != (B * S, P, C):
tokens = tokens.view(B, S, P, C).view(B * S, P, C)
if pos is not None and pos.shape != (B * S, P, 2):
pos = pos.view(B, S, P, 2).view(B * S, P, 2)
intermediates = []
# by default, self.aa_block_size=1, which processes one block at a time
for _ in range(self.aa_block_size):
tokens = self.frame_blocks[frame_idx](tokens, pos=pos)
frame_idx += 1
intermediates.append(tokens.view(B, S, P, C))
return tokens, frame_idx, intermediates
def _process_global_attention(self, tokens, B, S, P, C, global_idx, pos=None):
"""
Process global attention blocks. We keep tokens in shape (B, S*P, C).
"""
if tokens.shape != (B, S * P, C):
tokens = tokens.view(B, S, P, C).view(B, S * P, C)
if pos is not None and pos.shape != (B, S * P, 2):
pos = pos.view(B, S, P, 2).view(B, S * P, 2)
intermediates = []
# by default, self.aa_block_size=1, which processes one block at a time
for _ in range(self.aa_block_size):
tokens = self.global_blocks[global_idx](tokens, pos=pos)
global_idx += 1
intermediates.append(tokens.view(B, S, P, C))
return tokens, global_idx, intermediates
def slice_expand_and_flatten(token_tensor, B, S):
"""
Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing:
1) Uses the first position (index=0) for the first frame only
2) Uses the second position (index=1) for all remaining frames (S-1 frames)
3) Expands both to match batch size B
4) Concatenates to form (B, S, X, C) where each sequence has 1 first-position token
followed by (S-1) second-position tokens
5) Flattens to (B*S, X, C) for processing
Returns:
torch.Tensor: Processed tokens with shape (B*S, X, C)
"""
# Slice out the "query" tokens => shape (1, 1, ...)
query = token_tensor[:, 0:1, ...].expand(B, 1, *token_tensor.shape[2:])
# Slice out the "other" tokens => shape (1, S-1, ...)
others = token_tensor[:, 1:, ...].expand(B, S - 1, *token_tensor.shape[2:])
# Concatenate => shape (B, S, ...)
combined = torch.cat([query, others], dim=1)
# Finally flatten => shape (B*S, ...)
combined = combined.view(B * S, *combined.shape[2:])
return combined
vggt/vggt/models/vggt.py
+95
View File
@@ -0,0 +1,95 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn as nn
from huggingface_hub import PyTorchModelHubMixin # used for model hub
from .aggregator import Aggregator
from ..heads.camera_head import CameraHead
from ..heads.dpt_head import DPTHead
from ..heads.track_head import TrackHead
class VGGT(nn.Module, PyTorchModelHubMixin):
def __init__(self, img_size=518, patch_size=14, embed_dim=1024):
super().__init__()
self.aggregator = Aggregator(img_size=img_size, patch_size=patch_size, embed_dim=embed_dim)
self.camera_head = CameraHead(dim_in=2 * embed_dim)
self.point_head = DPTHead(dim_in=2 * embed_dim, output_dim=4, activation="inv_log", conf_activation="expp1")
self.depth_head = DPTHead(dim_in=2 * embed_dim, output_dim=2, activation="exp", conf_activation="expp1")
self.track_head = TrackHead(dim_in=2 * embed_dim, patch_size=patch_size)
def forward(
self,
images: torch.Tensor,
query_points: torch.Tensor = None,
):
"""
Forward pass of the VGGT model.
Args:
images (torch.Tensor): Input images with shape [S, 3, H, W] or [B, S, 3, H, W], in range [0, 1].
B: batch size, S: sequence length, 3: RGB channels, H: height, W: width
query_points (torch.Tensor, optional): Query points for tracking, in pixel coordinates.
Shape: [N, 2] or [B, N, 2], where N is the number of query points.
Default: None
Returns:
dict: A dictionary containing the following predictions:
- pose_enc (torch.Tensor): Camera pose encoding with shape [B, S, 9] (from the last iteration)
- depth (torch.Tensor): Predicted depth maps with shape [B, S, H, W, 1]
- depth_conf (torch.Tensor): Confidence scores for depth predictions with shape [B, S, H, W]
- world_points (torch.Tensor): 3D world coordinates for each pixel with shape [B, S, H, W, 3]
- world_points_conf (torch.Tensor): Confidence scores for world points with shape [B, S, H, W]
- images (torch.Tensor): Original input images, preserved for visualization
If query_points is provided, also includes:
- track (torch.Tensor): Point tracks with shape [B, S, N, 2] (from the last iteration), in pixel coordinates
- vis (torch.Tensor): Visibility scores for tracked points with shape [B, S, N]
- conf (torch.Tensor): Confidence scores for tracked points with shape [B, S, N]
"""
# If without batch dimension, add it
if len(images.shape) == 4:
images = images.unsqueeze(0)
if query_points is not None and len(query_points.shape) == 2:
query_points = query_points.unsqueeze(0)
aggregated_tokens_list, patch_start_idx = self.aggregator(images)
predictions = {}
with torch.cuda.amp.autocast(enabled=False):
if self.camera_head is not None:
pose_enc_list = self.camera_head(aggregated_tokens_list)
predictions["pose_enc"] = pose_enc_list[-1] # pose encoding of the last iteration
if self.depth_head is not None:
depth, depth_conf = self.depth_head(
aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
)
predictions["depth"] = depth
predictions["depth_conf"] = depth_conf
if self.point_head is not None:
pts3d, pts3d_conf = self.point_head(
aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx
)
predictions["world_points"] = pts3d
predictions["world_points_conf"] = pts3d_conf
if self.track_head is not None and query_points is not None:
track_list, vis, conf = self.track_head(
aggregated_tokens_list, images=images, patch_start_idx=patch_start_idx, query_points=query_points
)
predictions["track"] = track_list[-1] # track of the last iteration
predictions["vis"] = vis
predictions["conf"] = conf
predictions["images"] = images
return predictions
vggt/vggt/utils/geometry.py
+236
View File
@@ -0,0 +1,236 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import os
import torch
import numpy as np
def unproject_depth_map_to_point_map(
depth_map: np.ndarray, extrinsics_cam: np.ndarray, intrinsics_cam: np.ndarray
) -> np.ndarray:
"""
Unproject a batch of depth maps to 3D world coordinates.
Args:
depth_map (np.ndarray): Batch of depth maps of shape (S, H, W, 1) or (S, H, W)
extrinsics_cam (np.ndarray): Batch of camera extrinsic matrices of shape (S, 3, 4)
intrinsics_cam (np.ndarray): Batch of camera intrinsic matrices of shape (S, 3, 3)
Returns:
np.ndarray: Batch of 3D world coordinates of shape (S, H, W, 3)
"""
if isinstance(depth_map, torch.Tensor):
depth_map = depth_map.cpu().numpy()
if isinstance(extrinsics_cam, torch.Tensor):
extrinsics_cam = extrinsics_cam.cpu().numpy()
if isinstance(intrinsics_cam, torch.Tensor):
intrinsics_cam = intrinsics_cam.cpu().numpy()
world_points_list = []
for frame_idx in range(depth_map.shape[0]):
cur_world_points, _, _ = depth_to_world_coords_points(
depth_map[frame_idx].squeeze(-1), extrinsics_cam[frame_idx], intrinsics_cam[frame_idx]
)
world_points_list.append(cur_world_points)
world_points_array = np.stack(world_points_list, axis=0)
return world_points_array
def depth_to_world_coords_points(
depth_map: np.ndarray,
extrinsic: np.ndarray,
intrinsic: np.ndarray,
eps=1e-8,
) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
"""
Convert a depth map to world coordinates.
Args:
depth_map (np.ndarray): Depth map of shape (H, W).
intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
extrinsic (np.ndarray): Camera extrinsic matrix of shape (3, 4). OpenCV camera coordinate convention, cam from world.
Returns:
tuple[np.ndarray, np.ndarray]: World coordinates (H, W, 3) and valid depth mask (H, W).
"""
if depth_map is None:
return None, None, None
# Valid depth mask
point_mask = depth_map > eps
# Convert depth map to camera coordinates
cam_coords_points = depth_to_cam_coords_points(depth_map, intrinsic)
# Multiply with the inverse of extrinsic matrix to transform to world coordinates
# extrinsic_inv is 4x4 (note closed_form_inverse_OpenCV is batched, the output is (N, 4, 4))
cam_to_world_extrinsic = closed_form_inverse_se3(extrinsic[None])[0]
R_cam_to_world = cam_to_world_extrinsic[:3, :3]
t_cam_to_world = cam_to_world_extrinsic[:3, 3]
# Apply the rotation and translation to the camera coordinates
world_coords_points = np.dot(cam_coords_points, R_cam_to_world.T) + t_cam_to_world # HxWx3, 3x3 -> HxWx3
# world_coords_points = np.einsum("ij,hwj->hwi", R_cam_to_world, cam_coords_points) + t_cam_to_world
return world_coords_points, cam_coords_points, point_mask
def depth_to_cam_coords_points(depth_map: np.ndarray, intrinsic: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
"""
Convert a depth map to camera coordinates.
Args:
depth_map (np.ndarray): Depth map of shape (H, W).
intrinsic (np.ndarray): Camera intrinsic matrix of shape (3, 3).
Returns:
tuple[np.ndarray, np.ndarray]: Camera coordinates (H, W, 3)
"""
H, W = depth_map.shape
assert intrinsic.shape == (3, 3), "Intrinsic matrix must be 3x3"
assert intrinsic[0, 1] == 0 and intrinsic[1, 0] == 0, "Intrinsic matrix must have zero skew"
# Intrinsic parameters
fu, fv = intrinsic[0, 0], intrinsic[1, 1]
cu, cv = intrinsic[0, 2], intrinsic[1, 2]
# Generate grid of pixel coordinates
u, v = np.meshgrid(np.arange(W), np.arange(H))
# Unproject to camera coordinates
x_cam = (u - cu) * depth_map / fu
y_cam = (v - cv) * depth_map / fv
z_cam = depth_map
# Stack to form camera coordinates
cam_coords = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32)
return cam_coords
def closed_form_inverse_se3(se3, R=None, T=None):
"""
Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch.
If `R` and `T` are provided, they must correspond to the rotation and translation
components of `se3`. Otherwise, they will be extracted from `se3`.
Args:
se3: Nx4x4 or Nx3x4 array or tensor of SE3 matrices.
R (optional): Nx3x3 array or tensor of rotation matrices.
T (optional): Nx3x1 array or tensor of translation vectors.
Returns:
Inverted SE3 matrices with the same type and device as `se3`.
Shapes:
se3: (N, 4, 4)
R: (N, 3, 3)
T: (N, 3, 1)
"""
# Check if se3 is a numpy array or a torch tensor
is_numpy = isinstance(se3, np.ndarray)
# Validate shapes
if se3.shape[-2:] != (4, 4) and se3.shape[-2:] != (3, 4):
raise ValueError(f"se3 must be of shape (N,4,4), got {se3.shape}.")
# Extract R and T if not provided
if R is None:
R = se3[:, :3, :3] # (N,3,3)
if T is None:
T = se3[:, :3, 3:] # (N,3,1)
# Transpose R
if is_numpy:
# Compute the transpose of the rotation for NumPy
R_transposed = np.transpose(R, (0, 2, 1))
# -R^T t for NumPy
top_right = -np.matmul(R_transposed, T)
inverted_matrix = np.tile(np.eye(4), (len(R), 1, 1))
else:
R_transposed = R.permute(0, 2, 1) # (N,3,3)
top_right = -torch.bmm(R_transposed, T) # (N,3,1)
inverted_matrix = torch.eye(4, 4)[None].repeat(len(R), 1, 1)
inverted_matrix = inverted_matrix.to(R.dtype).to(R.device)
inverted_matrix[:, :3, :3] = R_transposed
inverted_matrix[:, :3, 3:] = top_right
return inverted_matrix
def depth_to_cam_coords_points_tensor(depth_map: torch.Tensor, intrinsic: torch.Tensor) -> torch.Tensor:
"""
Convert a depth map to camera coordinates.
Args:
depth_map (torch.Tensor): Depth map of shape (B, H, W).
intrinsic (torch.Tensor): Camera intrinsic matrix of shape (B, 3, 3).
Returns:
torch.Tensor: Camera coordinates (B, H, W, 3)
"""
B, H, W = depth_map.shape
# Intrinsic parameters
fu, fv = intrinsic[:, 0, 0], intrinsic[:, 1, 1]
cu, cv = intrinsic[:, 0, 2], intrinsic[:, 1, 2]
# Generate grid of pixel coordinates
v, u = torch.meshgrid(torch.arange(W).to(depth_map.device), torch.arange(H).to(depth_map.device))
# Unproject to camera coordinates
x_cam = (u[None] - cu[:, None, None]) * depth_map / fu[:, None, None]
y_cam = (v[None] - cv[:, None, None]) * depth_map / fv[:, None, None]
z_cam = depth_map
# Stack to form camera coordinates
cam_coords = torch.stack((x_cam, y_cam, z_cam), dim=-1).float()
return cam_coords
def depth_to_world_coords_points_tensor(
depth_map: torch.Tensor,
extrinsic: torch.Tensor,
intrinsic: torch.Tensor,
eps=1e-8,
) -> torch.Tensor:
"""
Convert a depth map to world coordinates.
Args:
depth_map (torch.Tensor): Depth map of shape (B, H, W, 1).
intrinsic (torch.Tensor): Camera intrinsic matrix of shape (B, 3, 3).
extrinsic (torch.Tensor): Camera extrinsic matrix of shape (B, 3, 4). OpenCV camera coordinate convention, cam from world.
Returns:
torch.Tensor: World coordinates (B, H, W, 3).
"""
if depth_map is None:
return None
# Valid depth mask
point_mask = depth_map > eps
# Convert depth map to camera coordinates
cam_coords_points = depth_to_cam_coords_points_tensor(depth_map, intrinsic)
# Multiply with the inverse of extrinsic matrix to transform to world coordinates
# extrinsic_inv is 4x4 (note closed_form_inverse_OpenCV is batched, the output is (N, 4, 4))
cam_to_world_extrinsic = closed_form_inverse_se3(extrinsic)
R_cam_to_world = cam_to_world_extrinsic[:, :3, :3]
t_cam_to_world = cam_to_world_extrinsic[:, :3, 3]
B, H, W, _ = cam_coords_points.shape
# Apply the rotation and translation to the camera coordinates
world_coords_points = torch.matmul(cam_coords_points.reshape(B, -1, 3), R_cam_to_world.float().permute(0,2,1)) + t_cam_to_world[:,None] # BxHWx3, Bx3x3 -> HxWx3
# world_coords_points = np.einsum("ij,hwj->hwi", R_cam_to_world, cam_coords_points) + t_cam_to_world
return world_coords_points.reshape(B, H, W, 3)
vggt/vggt/utils/load_fn.py
+118
View File
@@ -0,0 +1,118 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
from PIL import Image
from torchvision import transforms as TF
def load_and_preprocess_images(image_path_list):
"""
A quick start function to load and preprocess images for model input.
This assumes the images should have the same shape for easier batching, but our model can also work well with different shapes.
Args:
image_path_list (list): List of paths to image files
Returns:
torch.Tensor: Batched tensor of preprocessed images with shape (N, 3, H, W)
Raises:
ValueError: If the input list is empty
Notes:
- Images with different dimensions will be padded with white (value=1.0)
- A warning is printed when images have different shapes
- The function ensures width=518px while maintaining aspect ratio
- Height is adjusted to be divisible by 14 for compatibility with model requirements
"""
# Check for empty list
if len(image_path_list) == 0:
raise ValueError("At least 1 image is required")
images = []
alphas = []
shapes = set()
to_tensor = TF.ToTensor()
# First process all images and collect their shapes
# for image_path in image_path_list:
for img in image_path_list:
# Open image
# img = Image.open(image_path)
img = img[0]
# If there's an alpha channel, blend onto white background:
if img.mode == "RGBA":
# Create white background
alphas.append(to_tensor(img)[3:])
# background = Image.new("RGBA", img.size, (255, 255, 255, 255))
# Alpha composite onto the white background
# img = Image.alpha_composite(background, img)
# Now convert to "RGB" (this step assigns white for transparent areas)
img = img.convert("RGB")
width, height = img.size
new_width = 518
# Calculate height maintaining aspect ratio, divisible by 14
new_height = round(height * (new_width / width) / 14) * 14
# Resize with new dimensions (width, height)
img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)
img = to_tensor(img) # Convert to tensor (0, 1)
# Center crop height if it's larger than 518
if new_height > 518:
start_y = (new_height - 518) // 2
img = img[:, start_y : start_y + 518, :]
shapes.add((img.shape[1], img.shape[2]))
images.append(img)
# Check if we have different shapes
# In theory our model can also work well with different shapes
if len(shapes) > 1:
print(f"Warning: Found images with different shapes: {shapes}")
# Find maximum dimensions
max_height = max(shape[0] for shape in shapes)
max_width = max(shape[1] for shape in shapes)
# Pad images if necessary
padded_images = []
for img in images:
h_padding = max_height - img.shape[1]
w_padding = max_width - img.shape[2]
if h_padding > 0 or w_padding > 0:
pad_top = h_padding // 2
pad_bottom = h_padding - pad_top
pad_left = w_padding // 2
pad_right = w_padding - pad_left
img = torch.nn.functional.pad(
img, (pad_left, pad_right, pad_top, pad_bottom), mode="constant", value=1.0
)
padded_images.append(img)
images = padded_images
images = torch.stack(images) # concatenate images
alphas = torch.stack(alphas) # concatenate images
# Ensure correct shape when single image
if len(image_path_list) == 1:
# Verify shape is (1, C, H, W)
if images.dim() == 3:
images = images.unsqueeze(0)
alphas = alphas.unsqueeze(0)
return images, alphas
vggt/vggt/utils/pose_enc.py
+130
View File
@@ -0,0 +1,130 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
from .rotation import quat_to_mat, mat_to_quat
def extri_intri_to_pose_encoding(
extrinsics,
intrinsics,
image_size_hw=None, # e.g., (256, 512)
pose_encoding_type="absT_quaR_FoV",
):
"""Convert camera extrinsics and intrinsics to a compact pose encoding.
This function transforms camera parameters into a unified pose encoding format,
which can be used for various downstream tasks like pose prediction or representation.
Args:
extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4,
where B is batch size and S is sequence length.
In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world transformation.
The format is [R|t] where R is a 3x3 rotation matrix and t is a 3x1 translation vector.
intrinsics (torch.Tensor): Camera intrinsic parameters with shape BxSx3x3.
Defined in pixels, with format:
[[fx, 0, cx],
[0, fy, cy],
[0, 0, 1]]
where fx, fy are focal lengths and (cx, cy) is the principal point
image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
Required for computing field of view values. For example: (256, 512).
pose_encoding_type (str): Type of pose encoding to use. Currently only
supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
Returns:
torch.Tensor: Encoded camera pose parameters with shape BxSx9.
For "absT_quaR_FoV" type, the 9 dimensions are:
- [:3] = absolute translation vector T (3D)
- [3:7] = rotation as quaternion quat (4D)
- [7:] = field of view (2D)
"""
# extrinsics: BxSx3x4
# intrinsics: BxSx3x3
if pose_encoding_type == "absT_quaR_FoV":
R = extrinsics[:, :, :3, :3] # BxSx3x3
T = extrinsics[:, :, :3, 3] # BxSx3
quat = mat_to_quat(R)
# Note the order of h and w here
H, W = image_size_hw
fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
pose_encoding = torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
else:
raise NotImplementedError
return pose_encoding
def pose_encoding_to_extri_intri(
pose_encoding,
image_size_hw=None, # e.g., (256, 512)
pose_encoding_type="absT_quaR_FoV",
build_intrinsics=True,
):
"""Convert a pose encoding back to camera extrinsics and intrinsics.
This function performs the inverse operation of extri_intri_to_pose_encoding,
reconstructing the full camera parameters from the compact encoding.
Args:
pose_encoding (torch.Tensor): Encoded camera pose parameters with shape BxSx9,
where B is batch size and S is sequence length.
For "absT_quaR_FoV" type, the 9 dimensions are:
- [:3] = absolute translation vector T (3D)
- [3:7] = rotation as quaternion quat (4D)
- [7:] = field of view (2D)
image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
Required for reconstructing intrinsics from field of view values.
For example: (256, 512).
pose_encoding_type (str): Type of pose encoding used. Currently only
supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
build_intrinsics (bool): Whether to reconstruct the intrinsics matrix.
If False, only extrinsics are returned and intrinsics will be None.
Returns:
tuple: (extrinsics, intrinsics)
- extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4.
In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world
transformation. The format is [R|t] where R is a 3x3 rotation matrix and t is
a 3x1 translation vector.
- intrinsics (torch.Tensor or None): Camera intrinsic parameters with shape BxSx3x3,
or None if build_intrinsics is False. Defined in pixels, with format:
[[fx, 0, cx],
[0, fy, cy],
[0, 0, 1]]
where fx, fy are focal lengths and (cx, cy) is the principal point,
assumed to be at the center of the image (W/2, H/2).
"""
intrinsics = None
if pose_encoding_type == "absT_quaR_FoV":
T = pose_encoding[..., :3]
quat = pose_encoding[..., 3:7]
fov_h = pose_encoding[..., 7]
fov_w = pose_encoding[..., 8]
R = quat_to_mat(quat)
extrinsics = torch.cat([R, T[..., None]], dim=-1)
if build_intrinsics:
H, W = image_size_hw
fy = (H / 2.0) / torch.tan(fov_h / 2.0)
fx = (W / 2.0) / torch.tan(fov_w / 2.0)
intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device)
intrinsics[..., 0, 0] = fx
intrinsics[..., 1, 1] = fy
intrinsics[..., 0, 2] = W / 2
intrinsics[..., 1, 2] = H / 2
intrinsics[..., 2, 2] = 1.0 # Set the homogeneous coordinate to 1
else:
raise NotImplementedError
return extrinsics, intrinsics
vggt/vggt/utils/rotation.py
+138
View File
@@ -0,0 +1,138 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# Modified from PyTorch3D, https://github.com/facebookresearch/pytorch3d
import torch
import numpy as np
import torch.nn.functional as F
def quat_to_mat(quaternions: torch.Tensor) -> torch.Tensor:
"""
Quaternion Order: XYZW or say ijkr, scalar-last
Convert rotations given as quaternions to rotation matrices.
Args:
quaternions: quaternions with real part last,
as tensor of shape (..., 4).
Returns:
Rotation matrices as tensor of shape (..., 3, 3).
"""
i, j, k, r = torch.unbind(quaternions, -1)
# pyre-fixme[58]: `/` is not supported for operand types `float` and `Tensor`.
two_s = 2.0 / (quaternions * quaternions).sum(-1)
o = torch.stack(
(
1 - two_s * (j * j + k * k),
two_s * (i * j - k * r),
two_s * (i * k + j * r),
two_s * (i * j + k * r),
1 - two_s * (i * i + k * k),
two_s * (j * k - i * r),
two_s * (i * k - j * r),
two_s * (j * k + i * r),
1 - two_s * (i * i + j * j),
),
-1,
)
return o.reshape(quaternions.shape[:-1] + (3, 3))
def mat_to_quat(matrix: torch.Tensor) -> torch.Tensor:
"""
Convert rotations given as rotation matrices to quaternions.
Args:
matrix: Rotation matrices as tensor of shape (..., 3, 3).
Returns:
quaternions with real part last, as tensor of shape (..., 4).
Quaternion Order: XYZW or say ijkr, scalar-last
"""
if matrix.size(-1) != 3 or matrix.size(-2) != 3:
raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.")
batch_dim = matrix.shape[:-2]
m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(matrix.reshape(batch_dim + (9,)), dim=-1)
q_abs = _sqrt_positive_part(
torch.stack(
[
1.0 + m00 + m11 + m22,
1.0 + m00 - m11 - m22,
1.0 - m00 + m11 - m22,
1.0 - m00 - m11 + m22,
],
dim=-1,
)
)
# we produce the desired quaternion multiplied by each of r, i, j, k
quat_by_rijk = torch.stack(
[
# pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
# `int`.
torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),
# pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
# `int`.
torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),
# pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
# `int`.
torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),
# pyre-fixme[58]: `**` is not supported for operand types `Tensor` and
# `int`.
torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),
],
dim=-2,
)
# We floor here at 0.1 but the exact level is not important; if q_abs is small,
# the candidate won't be picked.
flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)
quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))
# if not for numerical problems, quat_candidates[i] should be same (up to a sign),
# forall i; we pick the best-conditioned one (with the largest denominator)
out = quat_candidates[F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :].reshape(batch_dim + (4,))
# Convert from rijk to ijkr
out = out[..., [1, 2, 3, 0]]
out = standardize_quaternion(out)
return out
def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:
"""
Returns torch.sqrt(torch.max(0, x))
but with a zero subgradient where x is 0.
"""
ret = torch.zeros_like(x)
positive_mask = x > 0
if torch.is_grad_enabled():
ret[positive_mask] = torch.sqrt(x[positive_mask])
else:
ret = torch.where(positive_mask, torch.sqrt(x), ret)
return ret
def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:
"""
Convert a unit quaternion to a standard form: one in which the real
part is non negative.
Args:
quaternions: Quaternions with real part last,
as tensor of shape (..., 4).
Returns:
Standardized quaternions as tensor of shape (..., 4).
"""
return torch.where(quaternions[..., 3:4] < 0, -quaternions, quaternions)
vggt/vggt/utils/visual_track.py
+239
View File
@@ -0,0 +1,239 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import cv2
import torch
import numpy as np
import os
def color_from_xy(x, y, W, H, cmap_name="hsv"):
"""
Map (x, y) -> color in (R, G, B).
1) Normalize x,y to [0,1].
2) Combine them into a single scalar c in [0,1].
3) Use matplotlib's colormap to convert c -> (R,G,B).
You can customize step 2, e.g., c = (x + y)/2, or some function of (x, y).
"""
import matplotlib.cm
import matplotlib.colors
x_norm = x / max(W - 1, 1)
y_norm = y / max(H - 1, 1)
# Simple combination:
c = (x_norm + y_norm) / 2.0
cmap = matplotlib.cm.get_cmap(cmap_name)
# cmap(c) -> (r,g,b,a) in [0,1]
rgba = cmap(c)
r, g, b = rgba[0], rgba[1], rgba[2]
return (r, g, b) # in [0,1], RGB order
def get_track_colors_by_position(tracks_b, vis_mask_b=None, image_width=None, image_height=None, cmap_name="hsv"):
"""
Given all tracks in one sample (b), compute a (N,3) array of RGB color values
in [0,255]. The color is determined by the (x,y) position in the first
visible frame for each track.
Args:
tracks_b: Tensor of shape (S, N, 2). (x,y) for each track in each frame.
vis_mask_b: (S, N) boolean mask; if None, assume all are visible.
image_width, image_height: used for normalizing (x, y).
cmap_name: for matplotlib (e.g., 'hsv', 'rainbow', 'jet').
Returns:
track_colors: np.ndarray of shape (N, 3), each row is (R,G,B) in [0,255].
"""
S, N, _ = tracks_b.shape
track_colors = np.zeros((N, 3), dtype=np.uint8)
if vis_mask_b is None:
# treat all as visible
vis_mask_b = torch.ones(S, N, dtype=torch.bool, device=tracks_b.device)
for i in range(N):
# Find first visible frame for track i
visible_frames = torch.where(vis_mask_b[:, i])[0]
if len(visible_frames) == 0:
# track is never visible; just assign black or something
track_colors[i] = (0, 0, 0)
continue
first_s = int(visible_frames[0].item())
# use that frame's (x,y)
x, y = tracks_b[first_s, i].tolist()
# map (x,y) -> (R,G,B) in [0,1]
r, g, b = color_from_xy(x, y, W=image_width, H=image_height, cmap_name=cmap_name)
# scale to [0,255]
r, g, b = int(r * 255), int(g * 255), int(b * 255)
track_colors[i] = (r, g, b)
return track_colors
def visualize_tracks_on_images(
images,
tracks,
track_vis_mask=None,
out_dir="track_visuals_concat_by_xy",
image_format="CHW", # "CHW" or "HWC"
normalize_mode="[0,1]",
cmap_name="hsv", # e.g. "hsv", "rainbow", "jet"
frames_per_row=4, # New parameter for grid layout
save_grid=True, # Flag to control whether to save the grid image
):
"""
Visualizes frames in a grid layout with specified frames per row.
Each track's color is determined by its (x,y) position
in the first visible frame (or frame 0 if always visible).
Finally convert the BGR result to RGB before saving.
Also saves each individual frame as a separate PNG file.
Args:
images: torch.Tensor (S, 3, H, W) if CHW or (S, H, W, 3) if HWC.
tracks: torch.Tensor (S, N, 2), last dim = (x, y).
track_vis_mask: torch.Tensor (S, N) or None.
out_dir: folder to save visualizations.
image_format: "CHW" or "HWC".
normalize_mode: "[0,1]", "[-1,1]", or None for direct raw -> 0..255
cmap_name: a matplotlib colormap name for color_from_xy.
frames_per_row: number of frames to display in each row of the grid.
save_grid: whether to save all frames in one grid image.
Returns:
None (saves images in out_dir).
"""
if len(tracks.shape) == 4:
tracks = tracks.squeeze(0)
images = images.squeeze(0)
if track_vis_mask is not None:
track_vis_mask = track_vis_mask.squeeze(0)
import matplotlib
matplotlib.use("Agg") # for non-interactive (optional)
os.makedirs(out_dir, exist_ok=True)
S = images.shape[0]
_, N, _ = tracks.shape # (S, N, 2)
# Move to CPU
images = images.cpu().clone()
tracks = tracks.cpu().clone()
if track_vis_mask is not None:
track_vis_mask = track_vis_mask.cpu().clone()
# Infer H, W from images shape
if image_format == "CHW":
# e.g. images[s].shape = (3, H, W)
H, W = images.shape[2], images.shape[3]
else:
# e.g. images[s].shape = (H, W, 3)
H, W = images.shape[1], images.shape[2]
# Pre-compute the color for each track i based on first visible position
track_colors_rgb = get_track_colors_by_position(
tracks, # shape (S, N, 2)
vis_mask_b=track_vis_mask if track_vis_mask is not None else None,
image_width=W,
image_height=H,
cmap_name=cmap_name,
)
# We'll accumulate each frame's drawn image in a list
frame_images = []
for s in range(S):
# shape => either (3, H, W) or (H, W, 3)
img = images[s]
# Convert to (H, W, 3)
if image_format == "CHW":
img = img.permute(1, 2, 0) # (H, W, 3)
# else "HWC", do nothing
img = img.numpy().astype(np.float32)
# Scale to [0,255] if needed
if normalize_mode == "[0,1]":
img = np.clip(img, 0, 1) * 255.0
elif normalize_mode == "[-1,1]":
img = (img + 1.0) * 0.5 * 255.0
img = np.clip(img, 0, 255.0)
# else no normalization
# Convert to uint8
img = img.astype(np.uint8)
# For drawing in OpenCV, convert to BGR
img_bgr = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
# Draw each visible track
cur_tracks = tracks[s] # shape (N, 2)
if track_vis_mask is not None:
valid_indices = torch.where(track_vis_mask[s])[0]
else:
valid_indices = range(N)
cur_tracks_np = cur_tracks.numpy()
for i in valid_indices:
x, y = cur_tracks_np[i]
pt = (int(round(x)), int(round(y)))
# track_colors_rgb[i] is (R,G,B). For OpenCV circle, we need BGR
R, G, B = track_colors_rgb[i]
color_bgr = (int(B), int(G), int(R))
cv2.circle(img_bgr, pt, radius=3, color=color_bgr, thickness=-1)
# Convert back to RGB for consistent final saving:
img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
# Save individual frame
frame_path = os.path.join(out_dir, f"frame_{s:04d}.png")
# Convert to BGR for OpenCV imwrite
frame_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
cv2.imwrite(frame_path, frame_bgr)
frame_images.append(img_rgb)
# Only create and save the grid image if save_grid is True
if save_grid:
# Calculate grid dimensions
num_rows = (S + frames_per_row - 1) // frames_per_row # Ceiling division
# Create a grid of images
grid_img = None
for row in range(num_rows):
start_idx = row * frames_per_row
end_idx = min(start_idx + frames_per_row, S)
# Concatenate this row horizontally
row_img = np.concatenate(frame_images[start_idx:end_idx], axis=1)
# If this row has fewer than frames_per_row images, pad with black
if end_idx - start_idx < frames_per_row:
padding_width = (frames_per_row - (end_idx - start_idx)) * W
padding = np.zeros((H, padding_width, 3), dtype=np.uint8)
row_img = np.concatenate([row_img, padding], axis=1)
# Add this row to the grid
if grid_img is None:
grid_img = row_img
else:
grid_img = np.concatenate([grid_img, row_img], axis=0)
out_path = os.path.join(out_dir, "tracks_grid.png")
# Convert back to BGR for OpenCV imwrite
grid_img_bgr = cv2.cvtColor(grid_img, cv2.COLOR_RGB2BGR)
cv2.imwrite(out_path, grid_img_bgr)
print(f"[INFO] Saved color-by-XY track visualization grid -> {out_path}")
print(f"[INFO] Saved {S} individual frames to {out_dir}/frame_*.png")