Closed RokuDoan closed 3 months ago
I have not noticed anything like that, results like that always happened with some seeds in my experience, unless you are comparing to exact same seed?
Can't really try same seed between update since I did not save the previous one. But retesting multiple seed did not improve this problem at all. But I pretty much nail consistent background output before this update.
Checked with the old commit git checkout 0ec07f6
and there definitely is a difference, but nothing as drastic (or even worse quality wise) than what you are experiencing, background is almost exactly the same:
One thing I just noticed when comparing: if the prompt is left empty, then the results can be drastically different, but with any prompt they become similar.
Edit: Okay I think the culprit is the clip vision model, only the original works well with some images, need to investigate further.
I think I found the problem, latest version should fix it.
Load CLIP model now causing this issue
Trying to load ToonCrafter model to get its CLIP now also get HeaderTooLarge
Load CLIP model now causing this issue
Trying to load ToonCrafter model to get its CLIP now also get HeaderTooLarge
I don't know what the first error is about, but on the 2nd you don't actually have the ToonCrafter model selected.
My bad, it looks like I tried to run the last workflow without noticing different model loaded in. The result seem to be pretty good now so I mark this one as closed.
After the recent update, it looks like the overall quality of animation seem to drop. Example below see a very minimal change in the background yet the generation seem to be frantic on its transition. On my testing 2 days before, it was a simply good side moving background transition.
https://github.com/kijai/ComfyUI-DynamiCrafterWrapper/assets/25916165/9acc58d6-e79d-4e65-9857-01c09a09290e
Even if there is no background movement, the character movement also become much less consistent.
Is it possible to revert to how it works before?