Repository navigation
Replies: 1 comment
|
Pretty much thats how Jasna works, latents from previous restorations are reinjected but you still have pretty large diffusion network to run where even 1 step takes a moment |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
The LTX model works by redrawing every frame from scratch. This redrawing strategy is computationally intensive by nature and results in relatively slow inference. I suggest the developers explore optimizations based on inter-frame comparison.
Most content that needs to be generated in a given frame overlaps heavily with the previous frame. It would be feasible to let each frame’s rendering reference the prior frame. Only the changed regions would then need redrawing. Meanwhile, the model should implement a threshold for cumulative error drift. This threshold limits the gradual buildup of visual artifacts caused by repeatedly referencing preceding frames. After each frame is generated, the model checks whether the error stays within the threshold. If the threshold is exceeded, the frame will be fully redrawn instead.
All reactions