Summary

CivitAI Anima LoRAs are trained with kohya-ss/sd-scripts, which writes keys as lora_unet_<munged_name>.lora_{down,up}.weight (+ optional .alpha), with dots in module paths replaced by underscores. The existing AnimaLoraLoaderMixin.lora_state_dict only handled the diffusion_model.* prefix with peft-style lora_A/lora_B suffixes, so kohya-format keys slipped through load_lora_weights silently and no PEFT adapter got registered. The user then sees a confusing Adapter name(s) {...} not in the list of present adapters later from set_adapters.

Changes

• _convert_kohya_anima_lora_to_diffusers (new) in lora_conversion_utils.py:
  • Pairs .lora_down.weight / .lora_up.weight / .alpha by base key.
  • Routes lora_unet_blocks_{i}_* to transformer.transformer_blocks.{i}.* (full self/cross attn, MLP layer1/layer2, adaLN modulation).
  • Routes lora_unet_llm_adapter_blocks_{i}_* to text_conditioner.blocks.{i}.*. Note: the LLM adapter uses o_proj (not the DiT's output_proj) and mlp.{0,2} (not the DiT's mlp.{layer1,layer2}). Verified against the live circlestone-labs/Anima/anima-base-v1.0.safetensors header.
  • Handles the non-block DiT leaves kohya excludes by default but some trainers include: x_embedder_proj_1, t_embedder_1_linear_{1,2}, final_layer_linear, final_layer_adaln_modulation_{1,2}.
  • Applies kohya alpha / rank scaling, distributing across down/up to avoid overflow (same formula as _convert_kohya_flux_lora_to_diffusers).
  • Skips lora_te_* Qwen3 text-encoder keys with a warning since the Anima loader does not load LoRAs into the text encoder.
• _convert_non_diffusers_anima_lora_to_diffusers (extended): also accepts kohya .lora_down/.lora_up/.alpha suffixes when the prefix is diffusion_model. (some trainers combine the two).
• AnimaLoraLoaderMixin.lora_state_dict: dispatches to the kohya converter when any key starts with lora_unet_ or lora_te_.

Test plan

• test_lora_state_dict_conversion (existing) — still passes.
• test_load_lora_weights (existing) — still passes.
• test_kohya_lora_state_dict_conversion (new) — covers DiT, LLM adapter, alpha scaling, and the TE-skip path.
• test_load_lora_weights_kohya_format (new) — end-to-end PEFT registration with kohya keys.
• Manual end-to-end against the real anima-base-v1.0.safetensors: 515/516 Linear modules round-trip cleanly; the 1 skipped is llm_adapter.embed (nn.Embedding), which kohya excludes by default anyway.

🤖 Generated with Claude Code