Did someone say... training? :doro: There's a working lora inside this discussion, trust me.
As description says, its possible to adapt diffusion-pipe to work with Anima, but needs a bit of tinkering, but a rank 32 Lora at 512px batch size 1 uses about 10GB with Qwen Frozen.
The bad news? diffusion-pipe only likes Linux
Here, have the lora 5e-5 AdamW, Batch 2, 512px with buckets
https://drive.google.com/file/d/1ec6YAjx14svK3grwhLfnVOwLrPU_4-zh/view?usp=sharing
Training code: https://github.com/bluvoll/diffusion-pipe runs as usual, added a dataset.toml, and anima.toml as examples.
Model uses 10gb of VRAM at batch size 1, 512px with Qwen Still loaded, gradient checkpointing and bf16
Previews:
NO LORA || LORA pairs.
I'm currently working on having shuffle and keep token working as similarly as it did in Kohya in a personal fork of diffusion-pipe.
More examples
oooor we could nicely ask tdrussell to update diffusion-pipe :P considering he's the creator of both diffusion-pipe and anima
Orrrr we could appreciate the nice gesture, since there’s no official training code, no LoRA support, and not even diffusers compatibility yet.
Besides, there's already a post asking for training code. https://huggingface.co/circlestone-labs/Anima/discussions/10
:3
oooor we could nicely ask tdrussell to update diffusion-pipe :P considering he's the creator of both diffusion-pipe and anima
My post is not meant to replace Tdrussell's code, but rather to help the community experiment with the model getting a hold of basic settings like Learning Rate, and VRAM usage, so we can use his code right out the gate.