nodes/bradsec-ComfyUI_ResolutionSelectorPlus
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix R1: per-model latent channel count (16 for Flux/Qwen/Z-Image); add Qwen-Image 4:3/3:4/3:2/2:3 + 1328 to Flux and Z-Image

Browse Source
This commit is contained in:
BRADSEC
2026-06-20 18:11:41 +10:00
parent e01cfac1ef
commit 9ac6653627
3 changed files with 73 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+3 -1
View File
@@ -8,9 +8,11 @@ Custom node for [ComfyUI](https://github.com/comfyanonymous/ComfyUI) providing e
### Resolution Selector
- **Model-optimised resolution presets** (portrait, landscape, square)
- **Empty latent output** for direct KSampler connection
- **Empty latent output** for direct KSampler connection, with the correct latent channel count per model (4 for SD 1.5 / SDXL, 16 for Flux / Qwen Image / Z-Image)
- **Custom width/height inputs** with **independent multiplier and latent**
Note: with model set to "All" the latent uses 4 channels. Pick the specific model (Flux, Qwen Image, Z-Image) to get a 16-channel latent for that model's sampler.
## Installation
```bash
resolution_selector.py
+44 -17
View File
@@ -11,34 +11,34 @@ except ImportError:
# Includes: model-optimized sizes + photo print (4x6, 5x7, 8x10) + digital/social + canvas art ratios
MODEL_RESOLUTIONS = {
"Flux": {
"square": [(512, 512), (768, 768), (1024, 1024), (1088, 1088), (1280, 1280), (1536, 1536), (1920, 1920), (2048, 2048)],
"portrait": [(688, 2048), (768, 1344), (832, 1216), (896, 1152), (928, 1664), (1024, 1536), (1024, 1792), (1024, 2048), (1088, 1920), (1152, 2048), (1200, 1792), (1360, 2048), (1456, 2048), (1536, 2048), (1616, 2048), (1632, 2048), (1712, 2048)],
"landscape": [(1280, 720), (1344, 768), (1216, 832), (1152, 896), (1536, 1024), (1664, 928), (1792, 1024), (1792, 1200), (1920, 1088), (2048, 688), (2048, 1024), (2048, 1152), (2048, 1360), (2048, 1456), (2048, 1536), (2048, 1616), (2048, 1632), (2048, 1712)],
"constraints": {"divisible_by": 16, "min": 256, "max": 2048}
"square": [(512, 512), (768, 768), (1024, 1024), (1088, 1088), (1280, 1280), (1328, 1328), (1536, 1536), (1920, 1920), (2048, 2048)],
"portrait": [(688, 2048), (768, 1344), (832, 1216), (896, 1152), (928, 1664), (1024, 1536), (1024, 1792), (1024, 2048), (1056, 1584), (1088, 1920), (1104, 1472), (1152, 2048), (1200, 1792), (1360, 2048), (1456, 2048), (1536, 2048), (1616, 2048), (1632, 2048), (1712, 2048)],
"landscape": [(1280, 720), (1344, 768), (1216, 832), (1152, 896), (1472, 1104), (1536, 1024), (1584, 1056), (1664, 928), (1792, 1024), (1792, 1200), (1920, 1088), (2048, 688), (2048, 1024), (2048, 1152), (2048, 1360), (2048, 1456), (2048, 1536), (2048, 1616), (2048, 1632), (2048, 1712)],
"constraints": {"divisible_by": 16, "min": 256, "max": 2048, "latent_channels": 16}
},
"Qwen Image": {
"square": [(1024, 1024), (1080, 1080), (1280, 1280), (1328, 1328), (1536, 1536), (1920, 1920), (2048, 2048)],
"portrait": [(680, 2048), (928, 1664), (1024, 1536), (1024, 2048), (1056, 1584), (1080, 1920), (1104, 1472), (1140, 1472), (1152, 2048), (1200, 1800), (1368, 2048), (1464, 2048), (1536, 2048), (1608, 2048), (1640, 2048), (1704, 2048)],
"landscape": [(1280, 720), (1472, 1104), (1472, 1140), (1536, 1024), (1584, 1056), (1664, 928), (1800, 1200), (1920, 1080), (2048, 680), (2048, 1024), (2048, 1152), (2048, 1368), (2048, 1464), (2048, 1536), (2048, 1608), (2048, 1640), (2048, 1704)],
"constraints": {"divisible_by": 8, "min": 256, "max": 2048}
"constraints": {"divisible_by": 8, "min": 256, "max": 2048, "latent_channels": 16}
},
"Z-Image": {
"square": [(512, 512), (768, 768), (1024, 1024), (1080, 1080), (1280, 1280), (1536, 1536), (1920, 1920), (2048, 2048)],
"portrait": [(680, 2048), (720, 1280), (768, 1024), (1024, 2048), (1080, 1920), (1152, 2048), (1200, 1800), (1368, 2048), (1464, 2048), (1536, 2048), (1608, 2048), (1640, 2048), (1704, 2048)],
"landscape": [(1024, 768), (1280, 720), (1800, 1200), (1920, 1080), (2048, 680), (2048, 1024), (2048, 1152), (2048, 1368), (2048, 1464), (2048, 1536), (2048, 1608), (2048, 1640), (2048, 1704)],
"constraints": {"divisible_by": 8, "min": 256, "max": 2048}
"square": [(512, 512), (768, 768), (1024, 1024), (1080, 1080), (1280, 1280), (1328, 1328), (1536, 1536), (1920, 1920), (2048, 2048)],
"portrait": [(680, 2048), (720, 1280), (768, 1024), (928, 1664), (1024, 2048), (1056, 1584), (1080, 1920), (1104, 1472), (1152, 2048), (1200, 1800), (1368, 2048), (1464, 2048), (1536, 2048), (1608, 2048), (1640, 2048), (1704, 2048)],
"landscape": [(1024, 768), (1280, 720), (1472, 1104), (1584, 1056), (1664, 928), (1800, 1200), (1920, 1080), (2048, 680), (2048, 1024), (2048, 1152), (2048, 1368), (2048, 1464), (2048, 1536), (2048, 1608), (2048, 1640), (2048, 1704)],
"constraints": {"divisible_by": 8, "min": 256, "max": 2048, "latent_channels": 16}
},
"SD 1.5": {
"square": [(512, 512), (768, 768), (1024, 1024), (1080, 1080), (1280, 1280), (1536, 1536)],
"portrait": [(512, 768), (512, 682), (512, 1024), (680, 2048), (768, 1024), (768, 1344), (1024, 2048), (1080, 1920), (1200, 1800), (1368, 2048), (1464, 2048), (1536, 2048), (1608, 2048), (1640, 2048), (1704, 2048)],
"landscape": [(768, 512), (1024, 512), (1024, 768), (1280, 720), (1344, 768), (1536, 512), (1800, 1200), (1920, 1080), (2048, 680), (2048, 1024), (2048, 1368), (2048, 1464), (2048, 1536), (2048, 1608), (2048, 1640), (2048, 1704)],
"constraints": {"divisible_by": 8, "min": 256, "max": 2048}
"constraints": {"divisible_by": 8, "min": 256, "max": 2048, "latent_channels": 4}
},
"SDXL": {
"square": [(1024, 1024), (1080, 1080), (1280, 1280), (1536, 1536), (1920, 1920), (2048, 2048)],
"portrait": [(640, 1536), (680, 2048), (768, 1344), (832, 1216), (896, 1152), (1024, 1536), (1024, 2048), (1080, 1920), (1152, 2048), (1200, 1800), (1368, 2048), (1464, 2048), (1536, 2048), (1608, 2048), (1640, 2048), (1704, 2048)],
"landscape": [(1152, 896), (1216, 832), (1280, 720), (1344, 768), (1536, 640), (1536, 1024), (1800, 1200), (1920, 1080), (2048, 680), (2048, 1024), (2048, 1152), (2048, 1368), (2048, 1464), (2048, 1536), (2048, 1608), (2048, 1640), (2048, 1704)],
"constraints": {"divisible_by": 8, "min": 256, "max": 2048}
"constraints": {"divisible_by": 8, "min": 256, "max": 2048, "latent_channels": 4}
}
}
@@ -184,6 +184,29 @@ def get_default_resolution(model_name):
return format_resolution(1024, 1024)
def get_latent_channels(model_name):
"""
Return the latent channel count for a model.
SD 1.5 and SDXL use 4-channel latents. Flux, Qwen Image, and Z-Image use
16-channel latents. A wrong channel count makes the LATENT output unusable
with that model's sampler.
For "All" (no specific model) this defaults to 4 for backward compatibility
with SD-based workflows; pick the specific model to get a 16-channel latent.
Args:
model_name (str): Model name, or "All".
Returns:
int: Latent channel count (4 or 16).
"""
model_data = MODEL_RESOLUTIONS.get(model_name)
if model_data:
return model_data["constraints"].get("latent_channels", 4)
return 4
def get_all_resolutions():
"""
Get all unique resolution strings across all models, sorted by dimensions.
@@ -346,8 +369,11 @@ class ResolutionSelector:
width *= multiplier
height *= multiplier
# Channel count depends on the model (4 for SD1.5/SDXL, 16 for Flux/Qwen/Z-Image)
channels = get_latent_channels(model)
# Generate latent for preset resolution with batch size
latent = self._generate_empty_latent(width, height, batch_size)
latent = self._generate_empty_latent(width, height, batch_size, channels)
# Determine final custom dimensions
if custom_width > 0 and custom_height > 0:
@@ -360,12 +386,12 @@ class ResolutionSelector:
self._validate_dimensions(model, final_custom_width, final_custom_height)
# Generate custom latent with custom batch size
custom_latent = self._generate_empty_latent(final_custom_width, final_custom_height, custom_batch)
custom_latent = self._generate_empty_latent(final_custom_width, final_custom_height, custom_batch, channels)
return (width, height, latent, final_custom_width, final_custom_height, custom_latent)
else:
# No custom dimensions, return zeros and minimal empty custom latent
custom_latent = self._generate_empty_latent(1, 1, 1)
custom_latent = self._generate_empty_latent(1, 1, 1, channels)
return (width, height, latent, 0, 0, custom_latent)
@@ -413,7 +439,7 @@ class ResolutionSelector:
f"{model} requires height between {min_dim} and {max_dim}. Got {height}"
)
def _generate_empty_latent(self, width, height, batch_size=1):
def _generate_empty_latent(self, width, height, batch_size=1, channels=4):
"""
Generate empty latent tensor for VAE input.
@@ -421,6 +447,7 @@ class ResolutionSelector:
width (int): Image width in pixels
height (int): Image height in pixels
batch_size (int): Number of latent samples (default: 1)
channels (int): Latent channel count (4 for SD1.5/SDXL, 16 for Flux/Qwen/Z-Image)
Returns:
dict: LATENT dict with 'samples' key containing empty tensor
@@ -429,9 +456,9 @@ class ResolutionSelector:
latent_width = width // 8
latent_height = height // 8
# Shape: [batch_size, channels=4, height//8, width//8]
# Shape: [batch_size, channels, height//8, width//8]
latent_tensor = torch.zeros(
[batch_size, 4, latent_height, latent_width],
[batch_size, channels, latent_height, latent_width],
device=self.device
)
test_resolution_selector.py
+26
View File
@@ -21,6 +21,7 @@ from resolution_selector import (
format_resolution,
get_resolution_list,
get_all_resolutions,
get_latent_channels,
parse_resolution_string,
MODEL_RESOLUTIONS
)
@@ -124,6 +125,29 @@ def test_all_resolutions_unique():
assert len(dimensions) == len(unique_dimensions), "Should have no duplicate dimensions"
print(f" ✓ All {len(all_res)} resolutions are unique")
def test_latent_channels():
"""R1: latent channel count must match the model (4 SD-based, 16 Flux/Qwen/Z-Image)"""
print("\nTesting latent channel mapping:")
expected = {"Flux": 16, "Qwen Image": 16, "Z-Image": 16, "SD 1.5": 4, "SDXL": 4, "All": 4}
for model, channels in expected.items():
got = get_latent_channels(model)
assert got == channels, f"{model}: expected {channels} channels, got {got}"
print(f" ✓ Channel mapping correct: {expected}")
def test_qwen_official_resolutions():
"""Official Qwen-Image sizes must be present in all 16-channel models"""
print("\nTesting official Qwen resolutions in 16-channel models:")
want = [(1328, 1328), (1664, 928), (928, 1664), (1472, 1104),
(1104, 1472), (1584, 1056), (1056, 1584)]
for model in ["Flux", "Qwen Image", "Z-Image"]:
resolutions = get_resolution_list(model)
for w, h in want:
s = format_resolution(w, h)
assert s in resolutions, f"{model} missing {w}x{h}"
print(f" ✓ All {len(want)} resolutions present in Flux, Qwen Image, Z-Image")
if __name__ == "__main__":
print("=" * 60)
print("ResolutionSelector Enhancement Tests")
@@ -137,6 +161,8 @@ if __name__ == "__main__":
test_model_resolutions()
test_new_resolutions()
test_all_resolutions_unique()
test_latent_channels()
test_qwen_official_resolutions()
print("\n" + "=" * 60)
print("✓ ALL TESTS PASSED!")