Beta 6 performs worse than Beta 5 for me – Could it be my ComfyUI workflow?
I've been having some issues with Beta 6. Beta 5 works significantly better for me, especially when it comes to prompt following and motion accuracy. Beta 6 looks slightly more realistic, but it struggles to follow my prompts properly.
Since many people are saying that Beta 6 is a big improvement over Beta 5, I'm wondering if the problem might be my ComfyUI workflow or settings.
I'd really like to try different workflows and see if I can get better results.
[I am using I2V and not T2V]
Could someone share a working Beta 6 ComfyUI workflow (.json)?
might be your prompting, i am using ref2v prompt template for every single one of my videos, and it works just well, i am using the Pixxorama workflow, with the recomended settings video shift 10 audio 3
I have a 2 stage beta 6 workflow with latent upscaling on my repository that works pretty nice.
Straight 8-10 steps single pass workflows are still preferred, but a two staged one is usually faster, and allows for quick previews.
https://huggingface.co/LokkenJP/TenEros_Max_optimized_learned if you want to take a look, in the SampleWorkFlowTwoPassOptimized subfolder
staying on beta5.
for me beta6 has terrible adherence to the provided image references.
staying on beta5.
for me beta6 has terrible adherence to the provided image references.
Same here. I've done several ref2video tests, and while beta6 seems to have better image quality, movements using character sheet references feel stiff. I think beta5 is much better in that regard.
Most should be taking my word and the comments more seriously and looking at their sampling, workflow and prompt. You have an issue there and I'm gonna start pushing back on model blamers. "Adherence", which is extremely broad by the way, can be a sampling shift cause or undersampling. b6 runs with 10/3 shift, 10 video shift while the normal model and b5 run 12. That's not suggested, it's required. It is literally built with full reference timestep and even has viggle in it. I've done video replacement, continuation, audio, music video lipsync, mixed multiple video references together, i2v done with reference node, 8 image reference composed outputs all perfectly fine and way beyond what 5 could do which is more or less just fl2va level.