Wan 2.2 v1 (T2V)
This is incredible. I did make a few changes to the dataset, added a few variety pieces, and I trained on 720p this time, but... I think the improvement is all Wan 2.2. Can't be certain, but... I'll let the samples speak for themselves.
Also, thanks to K3NK for the ridiculous workflow!
v1.1 (T2V)
This one is absolutely a one-trick pony, but that's kinda the point, right?
All of my videos always save the ComfyUI metadata, so right-click / drag-and-drop for a starting point!
v1.1 (I2V)
Same dataset as the T2V model, just trained on the 480P 14B model.
I2V seems to like higher shift, 12-15
it also seems to like higher CFG, like 6
oh, and a lower model weight on the LoRA, like 0.8
Experiment, and let me know / share your results!
a.k.a "The Adriano Angle"... IYKYK.
Post what you've got with this, I'd love to see some other successful prompts!
My first LoRA upload... This is WIP as I've been learning, but I figured I'd share what I got for feedback.
For images, I use the absolutely fantastic TAME Pony, which I don't think gets anywhere near the credit it deserves.
I followed dtwr434's very helpful Training Data guides for training WAN loras with diffusion-pipe, in fact I don't think I changed any settings at all for this one.
Description
Trimmed down the dataset a bit, re-captioned with a much better approach, and trained for 100 epochs. This one works WAY better T2V.
FAQ
Comments (10)
Cheers mate. Thanks for sharing! And I totally agree: "For images, I use the absolutely fantastic TAME Pony, which I don't think gets anywhere near the credit it deserves.
Fantastic work!! Congrats and thanks!
Sure, I can take a crack at that. Let me see what I can do.
@dngstn32 Great! Thank you very much!
@dngstn32 Hello! Sorry for my anxiety. Do you have any news about the penis and pussy standard version wan 1.3b t2v?
@rafaaguiar990170 Haven't had much luck yet. The settings I've tweaked so far seem to work for the 14B model, but with the 1.3b it doesn't seem to have much of a positive effect. I'm going to run another bunch of epochs soon, hopefully that improves things. If not, back to the drawing board!
@dngstn32 Ok, I see, this seems like a difficult task, I didn't know that generating penis and pussy would be so difficult for wan 1.3b t2v. I've never seen anyone manage to do this consistently yet. I'm really rooting for you to get a good result, this would be a great achievement for wan 1.3b t2v users! Thank you very much for your work and effort!! We'll be waiting for news!
✨ From my testing of the t2v model these settings worked really well:
Steps: 20
Model Sampling Shift: 6
CFG: 3.5
One thing I haven't done yet is play around with the shift settings on my T2V workflow, thanks for the pointer! If you're willing, please share some of your generations.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.