Using the wider FPS range for interpolation

[ambocclusion]

First of all, thank you for this amazing toolset!

One experiment I've been trying to do and failing somewhat is to use the new longer FPS range to interpolate shorter videos. If I generate longer videos I get a pretty chaotic scene, like so:
AnimateDiff_00194_.webm

I experimented with using ControlNet to use a generated shorter video to stabilize the scene and it seemed to have okay results:
AnimateDiff-interpolated_00004_.webm

Here is how I achieved that:

However, ComfyUI doesn't have tools to double up frames to fill a larger batch size, so all of the frames are just looped to fill the space. My question is: is it possible within ComfyUI to build a workflow to use AnimateDiff's longer video sizes to interpolate a shorter AnimateDiff video?

[AustinMroz]

Have you tried setting the context_stride in AnimateDiff Loader (Advanced) to a value greater than 1? Having a context_stride of 2 works very similarly to the method you describe here by performing an additional conditioning pass (in a single KSampler node) that skips every other frame.

[ambocclusion]

I have and it seems like it does seem to stabilize the output, it still changes context more than I would like. This is at 4:

AnimateDiff_00200_.webm

and here is at 2:

AnimateDiff_00202_.webm

[AustinMroz]

I appreciate the examples.

I don't think there's an existing way to repeat a batch like you described, but I could probably look into making a standalone custom node for testing with.

The more correct solution probably lies in new context schedulers that either put more bias on or perform additional passes on the higher strides, but it's not an area I personally plan to mess with any time soon. I feel I've caused too much of a distraction in that area with my failed uniform_v2 changes.

[ambocclusion]

Yeah admittedly what I'm asking for sounds like a serious hack and not a legitimate solution. I am just working on a project and looking for a solution. I may look into writing a custom node myself that can stretch a small amount of batches across a larger range, I just have to figure out where to start. Alternatively I can just hand-interpolate the videos and feed it all back through but the amount of videos I have to process is potentially large so it might be tedious.

I really appreciate the response!

[pondloso]

This is some of my experiment with new loader. It not pure t2video kind of half image2video and t2video
It hard to get consistently and creative at the same time.

i can do 10s or longer video with little denoise.
Only down side it take so freaking much time my sorry ass 3060TI just cry and beg me to stop.

AnimateDiff_00051_.mp4