V2V Extension: Continuous Color and Detail Drift
First, thanks for your contribution! Not sure if anyone else can relate, but i'm trying to stay consistent with 4k res as an end video leveraging the spatial upscaler (1.1 x2) and pass 1 frame by frame sources for color correction post process (pass 2 almost always 'burns' the pixels). I'd try to settle for lower res and post process upscale but in my case too much detail is lost if i do that. Particularly with V2V Extend use case, the more i extend the video (even with a 5090 i can't really do more than a few seconds at a time) the harder it is to keep the original color no matter how much i color correct with different methods individually or combined either in pixel or latent space. it's intended to be a long still scene.... and, i'd like to think i'm making a fixable mistake. Anyone come across and/or find a solution to this? thank u.
it's intended to be a long still scene
How long are you trying? LTX is trained for 20 sec max, and sweet spot probably around 10-15 sec.
You can extend lets say 15 seconds. Then use the extended video, and extend for yet another 15 sec... and so on.
and there are other ways such as the "long video" workflows that generates multiple extended videos in one batch (each part being say 10-20 seconds)
15 seconds at 1080p sound doable with 5090 + 96gb ram? Same scene in view about an hour (could split the hour w/ different clips) but, specific string of actions in the scene, about 15 seconds (the part that's tougher to split). IC Lora pose/motion has helped a lot for control with extension workflow. Without IC Lora, not so much. Taking a look at your long video workflow, you think using that template with, pehaps, IC lora guides for each section could work?