HairPulling DoggyStyle

Both work on their own ;) First one is trained on more videos. Second on is trained on fewer videos, but with additional high resolution images. So both work on their own, but give different output. Since I couldn't decide with one I prefer I just use both in the same generation at 0.5 and it works pretty well ;P But of course you can only use one of them if you like its output more than the other one or the combined one ;)

@makiaevelio543 hah yes , I also felt that both combined had more variantions in positions and camera angles =)

I'm using this: https://github.com/kohya-ss/musubi-tuner

yes T2V. was there a 480p and 720p model for t2v? thought it's just split for I2V

haven't tried it. but I think some examples from others in the gallery seems to be i2V

this one is trained at 24 fps. So playing it at the usuall 16 fps will make it look like in slow motion. My examples are on 24 fps, too

hm... maybe the Lora strength is too strong. should be 1, or if you combine both set them to 0.5 each. Also I noticed that vegatations tends to get a bit fuzzy compared to inside of a building for example.

don't know. Never looked into that stuff

@Geewhizz just had a look. the options is greyed out for me. probably because it's naugty stuff ;P

I switched both to 14b T2V. Is it working now ?

Glad you like it =)

have you tried less frames? I think all examples are at 49 frames. I used a very basic built in comfyUI setup. nothing fancy. I typically generate at 520x320 or 720x480. So your resolution should be fine

Thx =) . I'd had to redo my dataset preperation and training both Loras again. I think I'm too lazy right now ^^'' . But yes, should be strange to combine with other fps Loras. Hadn't thought much about it before training it ;P The workflows you linked look interesting! Maybe I can redo this Lora while testing them ;)

training parameter with musubi tuner were:

--sdpa --mixed_precision bf16 --fp8_base --optimizer_type adamw8bit --learning_rate 2e-4 --gradient_checkpointing --max_data_loader_n_workers 2 --persistent_data_loader_workers --network_module networks.lora_wan --network_dim 32 --timestep_sampling shift --discrete_flow_shift 3.0

so, nothing special. Resoution was 228x128 and 61 frames, around 25 clips. But currently I'm training with 456x256 while using --blocks_to_swap to not get out of RAM.

Key is to have a good dataset. Often I collect way to many mid quality clips, then I sorting out just the best, and with that smaller dataset I get better results. With 'quality' I don't mean video quality, but having the exact action and pose you want. And keep them similar to each other. Having a lot of angles and too few clipts of each just confuses the model. So this dataset only has side view videos.

didn't realize it had to be 16 while training it back then... at least it has more fluent motion now ;)