YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
MOSS-Image character LoRAs
Character adapters for MOSS-Image,
trained with ai-toolkit on MOSS-Image-Base and run against either checkpoint.
Each LoRA is a rank-32 / alpha-32 delta over the diffusion transformer only. The text encoder is frozen, so the text encoder and VAE are shared with the base model and never need to be swapped.
| id | trigger | strength | files |
|---|---|---|---|
alya |
alya(anime) |
1.0 | alya/MOSS_Alya_p100.safetensors |
emilia |
emilia(anime) |
0.85 | emilia/MOSS_Emilia.safetensors |
frieren |
frieren(anime) |
1.0 | frieren/MOSS_Frieren.safetensors |
marin |
marin(anime) |
1.0 | marin/MOSS_Marin.safetensors |
loras.json is the machine-readable index the
MOSS-IMAGE-LoRA Space reads at
startup: id, trigger, file, default strength, a one-line note, and an example
prompt.
Sampling
The recipes these adapters shipped with:
| setting | value |
|---|---|
| sampler | Euler, 50 steps |
| CFG | 3.0 |
| shift | 5.0 (built into the model) |
| negative-branch skip layer | 22 |
| negative prompt | nsfw, worst quality, bad anatomy, watermark, jpeg artifact |
Write prompts as natural language with the trigger in front: alya(anime), <description>.
Do not repeat appearance tags the trigger already covers (hair, eyes, ears) and do not
prepend booru style-tag strings; the training captions are natural-language recaptions.
Key layout
Tensors are named diffusion_model.layers.<n>.proj.<q|k|v|o|ffn_up_gate|ffn_down>.lora_{A,B}.weight,
which maps directly onto the diffusers MOSSImageTransformer module paths. Layer 22 is
excluded on purpose: it is the block the sampler skips on the negative branch, so training
it would fight inference.
Licensing
Inference code: Apache-2.0 (OpenMOSS). Weights: character frames come from BangumiBase
clusters and remain the property of their rights holders. These are not official OpenMOSS
releases. Per-adapter notes live in each folder's README.md.