Turbo LoRA for step and CFG distillation. Use CFG 1 and 8-12 steps. Works well on other Anima checkpoints. You can decrease the LoRA strength a bit below 1 for more variety.Work in progress. Increases stability, anatomy, and style consistency. Details and artist style adherence are a bit worse than base.The ER-SDE sampler might be too noisy and fried for this lora, but it's prompt dependent. Plain Euler works well and will be more neutral and flat.If your images look too "slopped":<ul><li>reduce lora weight, can go as low as 0.7 and still maintain distillation</li><li>remove quality tags, they aren't needed as much since a negative prompt is built in via the distillation</li><li>add "anime coloring" to the prompt</li></ul>

Anima Turbo LoRA

anima-turbo-lora-v0.1.safetensors

v0.1

v0.2

Turbo LoRA trained on preview3. Use CFG 1 and 8-12 steps. Works well on other Anima checkpoints. You can decrease the LoRA strength a bit below 1 for more variety.Work in progress. Increases stability, anatomy, and style consistency. Details are a bit worse than base.

Files with SHA256: 68ed0aec6ff4ebc3add1180e191797adb5aa6b69dd8b0fc8aa9e680145f65aac