MOSS Image 5B for WanGP

Prepared checkpoints for WanGP, based on OpenMOSS MOSS-Image. Base and SFT generate images from natural-language prompts, tags or experimental JSON layouts. SFT is post-trained from Base on curated data. The 5B size describes the image transformer; the Qwen3.5-9B text encoder is separate.

Select MOSS Image → 5B → Base / SFT in WanGP. Downloads are handled automatically. Start at 1024 × 1024, 50 steps, guidance 3 and shift 5. Neither checkpoint is step-distilled. Prompt enhancement is optional and disabled by default.

Files

Component	BF16	INT8 ConvRot
Base transformer	moss_image_base_bf16.safetensors	moss_image_base_int8_convrot.safetensors
SFT transformer	moss_image_sft_bf16.safetensors	moss_image_sft_int8_convrot.safetensors
Text encoder	Qwen3.5-9B/qwen3_5_9b_text_bf16.safetensors	Qwen3.5-9B/qwen3_5_9b_text_int8_convrot.safetensors
VAE decoder	moss_image/flux2_small_decoder_bf16.safetensors	—

The Qwen3.5-9B/ directory also contains the text-encoder configuration, tokenizer and chat template. Keep this hierarchy when downloading files manually.

The Prompt Helper edits text and object boxes; the layout enhancer uses the selected canvas size. Describe the complete scene and quote exact visible wording. Layout boxes guide composition, and lettering or placement can vary. These releases support text-to-image generation, without reference-image or pixel-mask inputs. Upstream recommends around 1–2 megapixels; dimensions must be multiples of 16.

Checkpoint preparation

These files use WanGP's runtime layout and INT8 ConvRot loader. They are not an unchanged Diffusers pipeline export.

• Transformers retain BF16 parameters, with the derived rotary-frequency buffer reconstructed at runtime. INT8 ConvRot versions quantize 252 matrices per transformer; other parameters remain BF16 and required quantization scales remain FP32.
• The original frozen Qwen3.5-9B encoder retains its language-conditioning weights; its unused vision tower is omitted. Query/gate rows are stored in contiguous groups for WanGP loading. The INT8 version quantizes 248 matrices.
• The small FLUX.2 VAE contains the decoder used by MOSS. Its unused image encoder is omitted. Weights are stored in BF16; required normalization statistics retain FP32 precision.

The prepared BF16 and INT8 formats were loaded and profiled with WanGP, then checked with complete 50-step, 1024-pixel poster generations for Base and SFT. Tensor mapping, storage dtypes and quantization metadata were verified. Quantization can change composition or lettering; these checks do not establish identical quality for every prompt.

Sources and license

Credit to the OpenMOSS Team for MOSS-Image, the Qwen Team for Qwen3.5-9B, and Black Forest Labs for the FLUX.2 small autoencoder used by the released model.

• MOSS-Image Base, revision 5f40f07a304541840b6019d2d34d30f966756787.
• MOSS-Image SFT, revision daf706d2e1c7a6369c17b6d1658671aee6db1ebe.
• Reference implementation.

The released MOSS weights and inference code are under Apache 2.0. See NOTICE and source revisions for provenance. This repository packages the released components for WanGP; it does not imply upstream endorsement.