MOSS-Image 5B lands in ComfyUI with ready-to-place weights

Comfy-Org has packaged OpenMOSS’s MOSS-Image-Base and MOSS-Image-SFT checkpoints for ComfyUI’s standard model directories. The weights bundle removes the need to assemble the original Diffusers pipeline, giving developers a simpler route to local generation, fine-tuning, and node-based workflows.

MOSS-Image is a 5B-parameter diffusion transformer developed by OpenMOSS, a research group backed by Fudan University, the Shanghai Innovation Institute, and MOSI Intelligence. Its main feature is structured prompting: alongside natural language and booru-style tags, the model can accept JSON layouts that specify the positions of text and visual elements.

One bundle, three model directories

The repository provides bf16 and quantized int8 versions of both image checkpoints, plus the Qwen3.5 text encoder and FLUX.2 VAE required for end-to-end inference. The files follow this directory structure:

ComfyUI/
  models/
    diffusion_models/
      moss_image_base_bf16.safetensors
      moss_image_base_int8_convrot_attn.safetensors
      moss_image_sft_bf16.safetensors
      moss_image_sft_int8_convrot_attn.safetensors
    text_encoders/
      qwen3.5_9b_int8_convrot.safetensors
    vae/
      flux2-vae.safetensors

• Diffusion model: The 5B image backbone denoises latent representations into the requested composition.
• Text encoder: A 9B Qwen3.5 model converts prompts and structured instructions into conditioning data.
• VAE: The FLUX.2 autoencoder converts the final latent representation into an image.

The 9B text encoder is larger than the image backbone, creating a substantial memory footprint even though MOSS-Image is described as compact. The prequantized int8 files reduce storage and VRAM pressure without requiring developers to convert the weights locally. Actual memory use depends on resolution, workflow, precision, attention implementation, and CPU offloading.