MOSS Character Studio Reference
A reference-conditioned LoRA and ComfyUI workflow for a complete four-panel character card: portrait, adult anatomy reference, clothed front and clothed back. The selected release uses one joint canvas, one sampler and one VAE decode. Reference image information is spatially aligned to the portrait bbox, while four-panel JSON layout and shared image attention are preserved.
Download and run
- Final BF16 LoRA
- Character Studio reference workflow
- API workflow
- Node package and installation
- Required reference setup
- Operation walkthrough
Install MOSS Card Layout and Wardrobe through ComfyUI Manager, or use the Git
URL above. This reference release requires the bundled, checksum-verified core
extension on ComfyUI 0.43.0. Follow REFERENCE_SETUP.md once, then restart.
The original MOSS HF snapshot is reused; the node pack does not download another
base-model copy. Place the LoRA under ComfyUI/models/loras/moss/.
Choose a face and outfit from the preview libraries, adjust character/card settings, inspect the final JSON, and run. Personal images can be added through Add my own. The Privacy coverage switch controls the anatomy reference panel; changing it regenerates the full card.
| Setting | Selected value |
|---|---|
| Base | Original MOSS-Image-SFT, native HF loader |
| Reference format | ai-toolkit |
| LoRA strength / trigger | 1.0 / none |
| Steps / CFG | 25 / 3.0 |
| Sampler / scheduler | Euler / sgm_uniform |
| Model sampling | AuraFlow, shift 5 |
| Skip-layer guidance | Single layer 22 |
| Default canvas | 1536 × 864 |
| Anchor / framing | full_portrait_panel / auto_portrait |
| Panel text scoping | Enabled |
| Text self-attention debug masks | Disabled |
Use the supplied reference workflow. The separate 16-step text accelerator is
not part of this recipe. Automatic framing aligns face position/scale/roll and
may suppress headwear or locally correct unequal eye geometry. It uses a verified
MediaPipe landmark model stored in models/mediapipe.
Image sources and credits
Face reference images come from the AI Portrait Gallery by あぎ. Costume reference images come from the AI Costume Gallery. These galleries supplied the face/outfit references used to build the training examples and preview libraries. The four-panel images displayed in this release were generated with the supplied character-card workflow. Original gallery images retain their source terms and are not bundled in this model release. Thanks to the original gallery creators.
Release examples
These are original generated adult anatomy/design studies with privacy coverage
turned off. They use the same synthetic gallery reference (fdff1a18c78) across
three outfits (SPR-0048, SPR-0006, PTY-0005) and three seeds (42, 1729,
719) to demonstrate unified identity preservation across wardrobe changes. The
original images carry their executed ComfyUI API graphs.
The walkthrough records actual face/outfit selection, the coverage switch, generation, prompt inspection and an outfit revision. Generation waiting intervals are shortened in the edited recording; the generated results are unchanged.
Training and provenance
This LoRA was further fine-tuned from akhaliq/MOSS-Image-Edit-HQ armV5, on the original OpenMOSS-Team/MOSS-Image-SFT. The continuation used 11 synthetic identities, 22 approved target cards with paired coverage On/Off, and three input-reference variants: 66 reference/target pairs. Two dataset passes produced 66 optimizer updates /132 sample forwards and backwards. A separate held-out identity was excluded from training.
Reference-input augmentation used generated scenes and deterministic source-face alignment/blending; inference starts with a noise canvas and does not paste the reference face into the output. Training and inference use the same reference anchor and cross-modal mask contract. The selected file is the unchanged 66-update BF16 adapter, with 504 tensors and mixed projection ranks inherited from armV5.
Filename: MOSS-Character-Studio-Reference-BF16.safetensors
Bytes: 495514728
SHA-256: b5bdc6a969a312025ce50ebe5f94463d5ad86dc5e0f260b915a8bfb1c806d477
The release examples retain the requested panel roles and selected garments. Some results retain short neck sections, change facial details/expression, change accessories, or show uneven skin illumination. Inspect generated cards before using them as production character sheets.
Weights are Apache-2.0, retaining attribution to OpenMOSS and the armV5 author. This is an independent community continuation. The node package is MIT licensed; its bundled ComfyUI core patch is GPL-3.0. Original face/outfit gallery images are linked or downloaded from their source and are not redistributed in this model repository. See the source galleries' terms for their images.