feat: MiniMax H3 Reference to Video works text-only; with no picture a plain stand-in of the Width/Height shape feeds the size chain (found on a real GPU)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

app/ui/features/generate/inputs.js

@@ -18,6 +18,13 @@ export async function buildInputs({ project, scene, slots, expand, frames, hooks
   for (const slot of (slots || []).filter((s) => s.node === "FunPackShotCamera")) put(slot, "project", project.id || "");      // what the node keeps per project (the words focus infusion draws on), also for pipelines saved before it was asked for
   const source = rolesAt(slots, "assets.source_image")[0];
   if (source) put(source.slot, "media_id", project.generation_mode === "t2v" ? "" : scene.source_image || "");    // "from a prompt" starts every shot without a picture
+  const like = source && source.role.blank_like;
+  if (like && !raw[source.slot.id].media_id) {          // text-only: a plain picture of the size the pipeline asks for stands in
+    const target = (slots || []).find((s) => s.id === like.slot) || { inputs: {} };
+    const size = (name) => (raw[target.id] || {})[name] ?? (target.inputs || {})[name];
+    put(source.slot, "blank_width", Number(size(like.width)) || 0);
+    put(source.slot, "blank_height", Number(size(like.height)) || 0);
+  }
   const plan = sceneReferences(scene, project.fixed_references);
   const { overrides, unwired } = wireReferences(plan.ids, slots, source ? [source.slot.id] : [], plan.pins);
   for (const [id, fields] of Object.entries(overrides)) raw[id] = { ...raw[id], ...fields };

app/ui/features/generate/inputs.test.js

@@ -82,3 +82,13 @@ test("the shot camera node is told which project this is, whatever roles the sav
   const { inputs } = await buildInputs({ project: { id: "proj1", scenes: [], video: {} }, scene: { id: "s", text: "x", references: [] }, slots, expand: async ({ text }) => ({ text }) });
   assert.equal(inputs.shot_camera.project, "proj1");
 });
+
+test("text-only: a slot that can stand in for the picture is told the size of the slot that shapes it, and a picked picture is sent as is", async () => {
+  const like = { slot: "t", width: "target_width", height: "target_height" };
+  const s = [{ id: "src", roles: [{ at: "assets.source_image", input: "media_id", blank_like: like }] },
+    { id: "t", inputs: { target_width: 1344, target_height: 768 }, roles: [{ at: "project.video", input: "target_width" }] }];
+  const run = (project, scene) => buildInputs({ project, scene, slots: s, expand: async () => null });
+  assert.deepEqual((await run({ generation_mode: "t2v", video: { target_width: 640 } }, { source_image: "m1" })).inputs.src, { media_id: "", blank_width: 640, blank_height: 768 });
+  assert.deepEqual((await run({ generation_mode: "i2v" }, { source_image: "m1" })).inputs.src, { media_id: "m1" });
+  assert.deepEqual((await run({ generation_mode: "i2v" }, {})).inputs.src, { media_id: "", blank_width: 1344, blank_height: 768 });
+});

core/tests/test_shipped_nodes.py

@@ -81,7 +81,7 @@ FROZEN = {
         "model", "lora_name", "strength_model", "clip", "strength_clip",
     ],
     "FunPackVAELoader": ["vae_name", "dtype"],
-    "FunPackLoadMedia": ["media_id"],
+    "FunPackLoadMedia": ["media_id", "blank_width", "blank_height"],
 }
 
 

modules/media/load/nodes.py

@@ -10,6 +10,7 @@ a run queues, the same way a scene's prompt text lands on CLIPTextEncode's
 `text`), not from someone typing a file name into this node by hand.
 """
 
+import torch
 from comfy_api.latest import InputImpl, io
 
 from ..._core import log, media as media_mod
@@ -27,13 +28,23 @@ class FunPackLoadMedia(io.ComfyNode):
                 io.String.Input(
                     "media_id", default="",
                     tooltip="A media library id, picked in the Media bin."),
+                io.Int.Input(
+                    "blank_width", default=0, min=0, max=16384, optional=True,
+                    tooltip="With nothing picked, make a plain grey picture this wide instead of refusing. 0 = refuse."),
+                io.Int.Input(
+                    "blank_height", default=0, min=0, max=16384, optional=True,
+                    tooltip="With nothing picked, make a plain grey picture this tall instead of refusing. 0 = refuse."),
             ],
             outputs=[io.Image.Output(display_name="image")],
         )
 
     @classmethod
-    def execute(cls, media_id: str) -> io.NodeOutput:
+    def execute(cls, media_id: str, blank_width: int = 0, blank_height: int = 0) -> io.NodeOutput:
         mid = (media_id or "").strip()
+        if not mid and blank_width > 0 and blank_height > 0:
+            # A text-only start: the picture exists only to give the pipeline a SHAPE (see the H3 preset).
+            log.info("FunPack Load Media", f"nothing picked: a plain {blank_width}x{blank_height} picture stands in")
+            return io.NodeOutput(torch.full((1, blank_height, blank_width, 3), 0.5))
         if not mid:
             raise RuntimeError(
                 "FunPack Load Media: nothing is picked for this input. "

modules/media/load/tests/test_load_media.py

@@ -59,3 +59,14 @@ def test_deleted_media_is_refused_even_with_a_well_formed_id(store):
     media.delete(entry["id"])
     with pytest.raises(RuntimeError, match="not in the library"):
         FunPackLoadMedia.execute(entry["id"]).result
+
+
+def test_a_blank_size_stands_in_for_nothing_picked_and_a_picked_image_still_wins(store):
+    from modules.media.load.nodes import FunPackLoadMedia
+    images, = FunPackLoadMedia.execute("", blank_width=64, blank_height=32).result
+    assert tuple(images.shape) == (1, 32, 64, 3)
+    entry = media.save_upload("ref.png", _png_bytes())
+    images, = FunPackLoadMedia.execute(entry["id"], blank_width=64, blank_height=32).result
+    assert images.shape[1:3] == (6, 4)
+    with pytest.raises(RuntimeError, match="nothing is picked"):
+        FunPackLoadMedia.execute("", blank_width=64, blank_height=0).result

modules/models/minimax_h3/pipeline.py

@@ -61,8 +61,11 @@ def h3_reference_to_video():
          # role's name and sets `media_id` from the scene's own picked image,
          # the same way "generation.prompt" is found by name and set from the
          # scene's own text -- neither is addressed by this preset's slot id.
-         "roles": [{"at": "assets.source_image", "input": "media_id"}],
-         "inputs": {"media_id": ""}},
+         # `blank_like`: with no picture (text-only), buildInputs gives this node the size of that
+         # slot's target_width/height, so a plain grey stand-in carries the project's own shape.
+         "roles": [{"at": "assets.source_image", "input": "media_id",
+                    "blank_like": {"slot": "image_transform", "width": "target_width", "height": "target_height"}}],
+         "inputs": {"media_id": "", "blank_width": 0, "blank_height": 0}},
         {"id": "image_transform", "group": "Reference media", "node": "ImageTransformKJ",
          # target_width/height are the PIXEL BUDGET, set in Project settings
          # like any other project resolution -- not the dropped picture's own