Prompt ComfyUI image VS Offical demo
Thanks to ComfyUI for supporting Lumina-Image 2.0. We have noticed that the images generated by ComfyUI differ significantly from those produced by our official demo, as shown below:
ComfyUI image:
Offical demo
The potential reasons for the differences are:
- CFG Renorm is vital for achieving accurate generation.
- System prompt is missing, which plays an important role in guiding the model.
- CFG truncation is not implemented, acceleration without compromising quality
These factors contribute significantly to the variations observed between ComfyUI and our official demo outputs.
if you use the same settings as the demo it matches.
ComfyUI uses torch.randn(device="cpu") for the noise while the demo uses cuda.
The 18 steps midpoint sampler in the demo is roughly equivalent to using 36 steps euler so make sure to do that and manually add the system prompt just like the example.
Whats the reasoning behind making it "all-in-one" file?
ComfyUI seemed to be made to allow custom workflows and modifying whatever is needed, not pre baked solutions.
And yea, creating noise on CPU vs GPU is quite important as basically anyone who used random noise from inspire pack knows. Not hard to change anyway.