Support for 5D latents (video models) and scheduling#62
Conversation
|
What's the status for this PR? It's been pending for months now, and @shiimizu seems to have given it a thumbs up. Edit: I've checked out this branch locally, everything seems in order. 👌 |
No problem, glad it ended up being helpful for you. I'd actually planned to give up and just close this but I forgot about it. Kind of doubt if there hasn't been any response after this much time that it's ever going to get merged. |
Please don't. I find it surprising that this custom node sees so little support in spite of its powerful capabilities and speed (relative to Ultimate SD Upscale which is extremely slow and clunky). This has been my go-to node for large resolution upscaling. |
|
I've run a number of inferences with this update on a Wan 2.2 t2i workflow, and the difference in quality is absolutely striking. |
|
Hi, thanks for taking interest in this PR, and thank you to @blepping for your work. Sorry for the late response. I do plan on merging it. It's just that there are other changes within, such as the |
@shiimizu Do you just mean remove the advanced parameters node? |
Allow scheduling the Tiled Diffusion effect Other minor cleanups
1169db0 to
265d5a4
Compare
3 participants
The main thing this pull adds is support for 5D latents so Tiled Diffusion can be used with video models and the ability to schedule the Tiled Diffusion effect.
Some refactoring was necessary to make applying settings function in a relatively sane way. Some of it, I just couldn't help myself. :) I tried to preserve the existing code style as much as possible.
One change I probably should explain is instead of having to track existing TiledDiffusion instances with a WeakSet, one can just keep a sequence number and update it whenever the instances need to reset. That way, instances can just save the sequence number from the last reset (or when they're created) and know to reset themselves when it changes.
These changes are lightly tested, as far as I can tell they don't break existing functionality. There is one case where they would break existing workflows: when a user is applying SpotDiffusion parameters before the TiledDiffusion model patch is applied. These changes require applying parameters after TiledDiffusion.
This pull also adds a TiledDiffusion Advanced Parameters node that currently allows configuring the start/end sigmas for the Tiled Diffusion effect. Adding other options, like allowing the user to specify a list of bboxes would be possible in the future. It would be pretty nice to be able to use custom bboxes instead of a fixed-size grid.
I'd like to add the ability to do temporal tiling but I figured I would make that a separate pull to keep these changes relatively manageable.
Small example of what this can enable:
AnimateDiff_00161.mp4