this run on 480 q4.1 fine

Confirming it works great with 480p fp8 as well.

ive used it on 480 it works perfectly fine

I'm using with 480 fp8 too. Works like a charm! Thanks

anyone got a workflow for i2v?

@ai_build_art Appreciate you sharing this info!

Wow, impressive you got this from 7 vids. I assume you had to use lower than the default [960, 544] resolution? (I can only go to 360x360x33 on my 3090.

@ai_build_art hey if you're up fpr it, I'd be willing to commission you if you to screen record your process? I'm more of a visual learner. Not even asking for a specific lora but moreso just understanding each step of the lora training process using video files, and how you prep/caption those files. Easily worth $100.

I'm very interested on know about the videos that you used for the training. Does these videos all has bouncing boobs ? Big boobs, saggy boobs ?

yes ! very good lora !!!!

Also two faces, you mean?

I agree. If it is too fast, Final Cut can slow it down and add frames. But if it is too slow, then speeding it up means a too-short clip that took way too long to render.

@ai_build_art I would also love to see a version trained at 16 fps! This LoRA currently does have a strong tendency to cause slow motion generations... Amazing LoRA though, thanks for sharing!

@gss3 Unfortunately I've discovered that Final Cut does not import 16 fps correctly. It just duplicates frames to make whatever the timeline is, and that makes the more advanced frame conforming and smooth slow motion not work.
The file needs to SAY that it is 24 frames per second, actually being that can be fixed after import.

@ai_build_art With the new CausVid or AccVid, CFG is 1, so no negative prompt :( And those already seem to make motion slower. They're becoming popular, so maybe next version can be trained with fast jiggly boobs, lol

@Algr I suggest using https://github.com/kijai/ComfyUI-GIMM-VFI to interpolate your clips before importing to Final Cut or Premiere Pro. But that's a separate problem from the one you raised initially with slow motion gens due to the loRA being trained on 30 fps videos (edit: which I think could only be truly solved by training at 16 fps)

@gss3  That is odd. If the training data is 30 fps, why lock output to 16 fps?
BTW, Both of those are unusual frame rates, all the high budget cinematography will be done at 24 fps, even when shot for American TV.

@Algr There are two separate but related issues here, and it's important to distinguish between them:

1. This LoRA was trained on 30 fps videos, while the base model (Wan 2.1) generates video at 16 fps. This mismatch causes problems. During LoRA training, the model learned from 30 frames for every second of video. That's almost twice the number of frames Wan 2.1 expects. From the base model's perspective, it's as if it was trained on slow-motion footage. So when you apply this LoRA during inference, it influences Wan 2.1 to generate output that also looks like slow-motion because the LoRA expects more motion to occur between frames than Wan 2.1 actually generates.

If the LoRA had been trained on 16 fps videos (to match the base model), this wouldn't be an issue. The motion and frame pacing would align, and you wouldn't see this slow-motion effect.

2. Wan 2.1 was trained to generate videos at 16 frames per second, and that's what it will always produce, regardless of what frame rate you want it to generate. This isn't a setting you can just tweak at inference time; it's baked into the model architecture and training data. During training, the engineers of Wan 2.1 chose 16 fps due to tradeoffs in prcessing power, memory, disk space, and training time. For example, training a model on 24 fps video might require up to 50% more compute and storage, which is a big deal when training on 1.5 billion videos.

As a result, any video generated by Wan 2.1 will always have 16 fps because that's all it knows how to do. You can't ask it to generate 24 fps directly, because the temporal resolution of the model is fixed by its training.

If you need 24 fps output, the correct workflow is to first generate your 16 fps video with Wan 2.1, and then interpolate it to 24 fps using a frame interpolation tool (or nodes to do this in ComfyUI). Interpolation works by analyzing the motion between frames and generating entirely new intermediate frames to smooth out the motion. It's not just duplicating existing frames. The project I linked above can do this. https://github.com/Fannovel16/ComfyUI-Frame-Interpolation is another good option for ComfyUI. Once you've interpolated your video to 24 fps, you can import it into Final Cut Pro and work with it just like any other 24 fps footage.

I hope this helps!

@gss3 we just need to train our loras at 6fps then to speed it up 🤣

@ai_build_art yes please retrain at 16fps. this would be perfect if trained at the correct frame rate. you don't need fancy tools, just load in comfy with vhs at 16fps, and save at 16fps. boom done with conversion.

@gss3 The actual images generated don't need to change. It just needs the quicktime or mp4 file to say that this is 24p and not 16p. Given that these files are mostly slow motion, this is actually closer to the visual truth anyway. Ultimately this is a bug in Final Cut, not WAN. If Final Cut could import 16p without generating extra frames in a dumb way, this would be unnecessary. Unfortunately, my experience is that big companies just don't fix bugs of this kind.

While you're right this would be better with more samples, sometimes collecting those is harder than it seems. Though it shouldn't be THAT hard. Also, Gemini wouldn't have any idea anyway, and would just make educated guesses.

@theft_of_life how did gemini get to that conclusion

Eh, the method seems to work, for images I can easily imagine walking and breasts bouncing.

@fitCorder halucination, AI almost never tell you I don't know the answer.
Also this 7 videos can be enough because training wan on lora that doesn't include genitals is very easy compared to the +18 content.

@Dead_Internet_Theory while porn is abundantly available, I've found when looking for a specific pose/position/direction it's surprisingly more difficult than one would expect. Never cared about that before, heh.

Having the same problem too now. I usually manually add resource when I post. With the change of "base model" category, I can't add the wan checkpoint and another wan lora because this bouncy and your bouncy walk lora are now "Wan Video 14B i2v 720p" instead of regular "WAN Video". It's bad enough (or because I'm very due diligence) that I split add my resource used in the picture before i2v AND after i2v. Now I have to add the resource used for the OG picture generated + i2v resources use under "WAN Video" + yours under "Wan Video 14B i2v 720p"

Is there a way for you to change this back to just "Wan Video"?

Edit, the following statement is not mad at you, I'm mad at Civitai: oh god, now that I researched a little more, theres FIVE wan categories now. So what if I used a wan checkpoint thats under "WAN Video" but now I can't add yours under Wan Video 14B i2v 720p. What if I use a 720p resource to make 480p compatible stuff? What about those mix i2v/t2v resource, who should get credit then for using a lora? What about old loras under WAN category that's totally usable but you can't use them now because author is gone and haven't update the tag to the new categories? This is gonna split the resources used to generate even more.

@ai_build_art I went to the Civitai suggestion page and posted about this problem. Its not admin approved yet but once it does, I hope anyone that hates this problem as much as me, go say/upvote on the thread or give your two cents.