Workflows with high fidelity
Hi I was testing a lot workflows online and many of them had a lot issues such as color shift, zoom instead of motion and even fixed image instead of video, but with this repo there's no issue whatsoever. So big thanks for that. My question is: are you planning to make any workflows with high fidelity? Current ones have cfg 1, so is there any chnace to see wf with cfg like 3.5, of course I'm aware it will take more steps to generate and low step lora need to be disabled/ on lower strength. Im mostly interested in i2v and as I see many people have issue with i2v in ltx
Happy to hear ;-)
Yes, planning to make HQ Dev workflow that has more steps, CFG higher etc. Its gonna be slow to run, but for those that want to squeeze out the little extra of the model ;-)
That being said, in the workflows here (i think most or all), there is a NAG node, that lets you use negative prompting even with CFG 1.
(but still, a full dev workflow can get even better results, and you can pump up steps etc)
Can't wait for it. Hope it's going to be released soon. With current i2v wf I struggle very often in selfie-like vertical videos I feel like 90% of videos I generated there's anomalies while person is moving her hands. Only while trying to generate many times some seeds were able to generate without it.