Works for WAN2.2 a14b t2v and i2v.
All example videos generated using both high & low lora at strength 1.0
All showcase videos have workflow embedded. (Use the download button to preserve it, if you Right-click/save-as the workflow data gets stripped by the browser)
Description
entirely new dataset with more varied breast sizes, ethnicities, body types
FAQ
Comments (6)
Hey there. Could you explain how you went from a 300mb V1 to 34mb V2 files? Thank you 🙌
Hey, sure thing! Two major contributors:
- The LoRA is dim 8. Previous version was likely dim 32 or 16, can't remember. File size is directly proportional to dim, doubling dim also doubles file size. As a 14b parameter model, WAN doesn't really need a large dim to pick up a concept. (Unlike typical image models like Klein 4b or 9b)
- The LoRA has had its blocks stripped to only affect 4-24 (out of 40) This further reduces the file size by the same proportion. For WAN, earlier blocks affect general image composition while later blocks affect micro details. As an anatomical LoRA, it only needs to affect early blocks and can let WAN use its base knowledge for final texture. From experimentation, blocks 25-40 can pretty much always be dropped and will make your LoRA better, because it will prevent it from getting that 'fried' look, and will actually behave better when combined with other LoRAs.
@2_as_1 Thank you, it makes sense. I only knew about the dim, so thank you very much for explaining the blocks, even if I have no idea how you do that. I only wish there were great training tutorials on CivitAI for Wan. Cheers 🙌
@GlowingGuardianGirl If you're after the benefits of stripped blocks, you can achieve that in ComfyUI by using a 'lora block loader' node instead of a regular lora loader node. That won't reduce the file size, but it can make certain LoRAs behave a little better.
Hello, thank for the model! i am having some trouble in I2V, for some reason almost the same prompt works with triboob but with quadboob not, are there any recomendations?
convet to ZImageBase please