video generation UI improvement

[dobkun]

DT currently reuses image generation toggles and thingamajigs for video models like LTX-2 which essentially prevents people from using model-exclusive toggles like FPS, image reference slot(instead of image as first frame) and more. by the way, this prevents LTX-2 from generating videos longer than 10 seconds even though it is technically possible and would have been a good thing to have for paid users. this and also confusing lingo like this:

[dobkun]

oh and by the way changing CFG(A PARAMETHER THAT DOES ONLY AFFECT LOGIC AND NOT RESOURCE CONSUMPTION) mysteriously drains more compute units

[KittenWizard]

TL;DR: Setting CFG to 1.0 uses roughly half the resource units and is quicker to inference.

CFG works by calculating the difference between the noise prediction of the positive prompt and the unconditional (or negative) prompt, then scaling that difference to influence the final image.

The underlying equation for this guidance is often represented as: N=U+W×(P−U) Where P is the positive prompt noise prediction, U is the unconditional noise prediction, and W is the CFG scaling value.

CFG requires two forward passes (one for positive, one for negative), making it computationally slower than having CFG disabled which is a CFG value of 1.0

Perhaps that is what you're noticing?