Jukka Seppänen

3D modeling/printing artist, AI enthusiast, rookie Python coder.

[Jukka Seppänen @Kijaidesign]

@Chris_Scott_Dev It's available currently in my WanVideoWrapper -nodes.

[Jukka Seppänen @Kijaidesign]

WanVideo2video fun, 129 frames with sliding context windows using the 14B model.

[Jukka Seppänen @Kijaidesign]

@SlipperyGem I2V is bit more complicated, haven't figured out best way to do that yet.

[Jukka Seppänen @Kijaidesign]

WanVideo sliding context window test on the 480p 14B T2V -model, worked surprisingly well but incredibly slow (50 mins on a 5090, 513 frames at 832x480)

[Jukka Seppänen @Kijaidesign]

@jqlive Absolutely, I have the rudimentary implementation in my HunyuanVideo wrapper nodes in ComfyUI as well, it needs some updating to be as good though.

[Jukka Seppänen @Kijaidesign]

@XCeciiiiiiiiiil This is with my ComfyUI wrapper nodes yeah.

[Jukka Seppänen @Kijaidesign]

@6___0 It's due to the technique itself, each step basically does the whole video in chunks, which include the overlaps, so many many passes though the model required. Actual VRAM use is really low since only the 81 frames are processed at once, as that's what the model does best.

[Jukka Seppänen @Kijaidesign]

@melmassadian @el_mejnun No way, my pride wouldn't take that, I've built every PC I've ever had myself since I was 14!

[Jukka Seppänen @Kijaidesign]

WanVideo has been lot of fun, currently playing with applying sliding context windowing similar to AnimateDiff, the 1.3B model is especially suitable for this due to it's speed, here's 1025 frames in one go, under 5GB VRAM used, but took about 10 mins on a 5090 with 30 steps.

[Jukka Seppänen @Kijaidesign]

As an impatient person, I often get obsessed with optimizations, tried implementing TeaCache for WanVideo, failed, but in the process accidentally (maybe) succeeded? Results look promising at least!

[Jukka Seppänen @Kijaidesign]

@zazoum1 Currently only single prompt with sliding context windows, but I plan to test multiple ones if I can figure all that out.

[Jukka Seppänen @Kijaidesign]

@gjohgj The model is supposed to output 16fps, though people seem to debate that. This video was saved with 16fps.

[Jukka Seppänen @Kijaidesign]

@el_mejnun I actually have both, I couldn't justify upgrading but 2nd rig for all this stuff I couldn't resist, so I splurged.

It's considerably faster, around ~40% or so with all optimizations. Slightly difficult to use still as you have to compile many things from source. Runs hot though

[Jukka Seppänen @Kijaidesign]

@YolaoDude Any text2video model can do some level of video2video as long as the VAE encoder is available, and it already was. The process is basically same as the usual img2img.

[Jukka Seppänen @Kijaidesign]

HunhuyanVideo -model's vid2vid passes the hippo test! Very promising and versatile model, thanks to the cfg distillation and every possible optimization I could think of this clip of 101 frames at 768x432 took about 2 minutes to sample, fitting slightly under 20GB using my nodes.

[Jukka Seppänen @Kijaidesign]

Couple more examples, just changing prompt.

[Jukka Seppänen @Kijaidesign]

@ZenMatAI oh and available here: https://t.co/Pf1MmKgsPO

[Jukka Seppänen @Kijaidesign]

@ZenMatAI It's the same model, but at fp8 so half the size.

[Jukka Seppänen @Kijaidesign]

@stilfuchs Oh wow that worked surprisingly good, this is on my list of things to explore as well! Thanks for sharing :D

[Jukka Seppänen @Kijaidesign]

I have finally pushed a bigger update to my CogVideoX ComfyUI wrapper nodes, cleaning up most of the bloat that has been accumulating with all these different models. One of the discoveries I made during this is that the orbit -LoRAs work with the "Fun" -models as well!

[Jukka Seppänen @Kijaidesign]

@OneStrangeW The workflow, among with many (much simpler) others are included with the nodes. I'm afraid I can't help with huggingface issues as all the models are hosted there.