MOSS-Image inference architecture adapted from OpenMOSS:
https://github.com/OpenMOSS/MOSS-Image
https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base
Base revision: 5f40f07a304541840b6019d2d34d30f966756787
SFT revision: daf706d2e1c7a6369c17b6d1658671aee6db1ebe
The released weights and inference code are licensed under Apache-2.0.

The frozen text encoder is Qwen3.5-9B, using WanGP's existing Qwen
implementation and its preserved upstream license notices. The unused
vision tower is omitted. Q/gate projection rows are stored in separate
contiguous groups in the WanGP encoder checkpoint.

The FLUX.2 small VAE reuses WanGP's FLUX autoencoder implementation. Text-to-image
uses the decoder-only checkpoint; reference editing uses the full encoder and
decoder checkpoint with the original normalization statistics.

MOSS Image Edit adapter by chinoll:
https://huggingface.co/chinoll/Moss-Image-Edit
Adapter revision: 3be9c40037e1b89b17aa85ff6f8ee805872aa558
The adapter is trained against MOSS Image SFT. Its rank/alpha configuration
and tensor values are preserved in the packaged adapter. The adapter repository
does not declare a license; the base model's Apache-2.0 notice does not establish
the license of this separately released adapter.

Reference-conditioning contract inspected in Chenkin-x/Moss-Image-Edit
(d735569c936ebbfaf274bee5cbff4297c533d133) and its ComfyUI integration
(b3377be990c0fa296459a20a4503dda40c045b4b). The WanGP implementation extends
its existing MOSS transformer and shared attention operations.

MOSS-Image-Edit-HQ v4 adapter by akhaliq:
https://huggingface.co/akhaliq/MOSS-Image-Edit-HQ
Revision: e67d778c32f20756dd9bd1fc2c1d0f06c3c34b42
Released file: armV4-moss-edit-lora.safetensors
The repository declares Apache-2.0. This is an unofficial community fine-tune
of MOSS Image SFT, not an official OpenMOSS release. The packaged adapter
preserves its BF16 tensors, with explicit alpha=rank (192 for q/k/v and 64
for output/FFN projections), matching the released training configuration.
HQ uses the ai-toolkit inference contract at revision 14f8f95:
https://github.com/ostris/ai-toolkit/tree/14f8f95/extensions_built_in/diffusion_models/moss_image
Clean references use timestep zero, indexed reference positions and the
current prompt length on the text axis. They attend only to references.
The negative branch skips layer 22, including its reference state, and
has its own reference KV cache. The original Edit SFT workflow is unchanged.
