About these variants
Four precision levels are available for MOSS-Image (available for both the SFT and Base checkpoints), trading VRAM and speed against generation quality:
- BF16 — full-precision reference (~10 GB). Highest quality, largest footprint. Available as diffusers on the official OpenMOSS-Team/MOSS-Image-SFT and OpenMOSS-Team/MOSS-Image-Base Hugging Face repositories, and ComfyUI compatible models on the Comfy-Org/MOSS-Image Hugging Face repository.
- INT8 (W8A8) ConvRot — 8-bit weights + 8-bit activations (~5.4 GB). Near-lossless quality (mean relative error ~0.87%, cosine similarity ~1.0005 vs BF16), ~2× smaller than BF16, runs on INT8 tensor cores (RTX 30-series and up).
- INT6 (W6A8) — 6-bit weights + 8-bit activations (~4.1 GB). Middle-ground: smaller than INT8 with only a small quality trade-off.
- INT4 (W4A8) — 4-bit weights + 8-bit activations (~3.0 GB). Smallest footprint; uses ConvRot with a per-tensor codebook that decodes to INT8 for compute, so it runs on the same INT8 hardware as W8A8. Larger quality trade-off than INT6.
Two checkpoints available at each precision, same architecture, different intended use:
- MOSS-Image-SFT (moss_image_sft_*) — supervised fine-tuned on top of Base. The default for day-to-day text-to-image generation, including natural-language prompts, tag sequences, and JSON layout control.
- MOSS-Image-Base (moss_image_base_*) — foundation checkpoint for continued pretraining, fine-tuning, and LoRA training. Use this as the starting point when training adapters or doing further post-training; SFT is the better choice for end-user inference.
MOSS-Image — INT8 / INT6 / INT4 ConvRot for ComfyUI
Quantized weights of MOSS-Image for fast, low-VRAM inference in ComfyUI. These are modified (quantized) versions of the official MOSS-Image checkpoints from the OpenMOSS Team, repackaged for ComfyUI with mixed-precision quant metadata that loads via the native MOSS-Image ComfyUI nodes. The BF16 reference is linked for comparison but is available directly from OpenMOSS Team's official repositories.
About MOSS-Image
A text-to-image diffusion model from the OpenMOSS Team, with a 5B DiT image transformer (42 layers, 252 GEMM linear projections) paired with a Qwen3.5-9B text encoder and a FLUX.2 small VAE. MOSS-Image supports natural-language prompts, tag-sequence prompts, and JSON-layout conditioning for precise text placement and object arrangement in the generated image.
Highlights
- Flexible Prompting — accepts free-form natural language, compact tag sequences, and structured JSON layouts in the same model, letting creators match the prompting style to the task.
- Layout-to-Image Control — pass a JSON layout with a scene caption, items, bounding boxes (yxyx format, abs_centralized coordinates), and rendered-text specifications to control where elements appear in the frame.
Key features
- Unified Prompting Interface — natural-language, tag-sequence, and layout-JSON prompts all handled by one model, no mode switch.
- Rendered Text and Layout — specify on-image text and element positions directly through JSON, useful for posters, signage, UI mockups, and composition-driven work.
- Qwen3.5-9B Text Encoder — large language-model-grade text understanding for stronger prompt grounding than CLIP-class encoders.
- FLUX.2 Small VAE — shared autoencoder path with the FLUX.2 ecosystem; VAE kept in BF16 across all quantized variants (quantizing a VAE this small saves little and risks reconstruction artifacts).
- SFT and Base Variants — SFT for inference, Base for fine-tuning and LoRA training.
Recommended inference settings (from the official MOSS-Image ComfyUI workflow):
- Sampler: Euler
- Scheduler: sgm_uniform
- Steps: 50
- CFG: 3
- Shift: 5
- Negative-branch skipping: layer 22
- Resolution: 1024×1024
License
MOSS-Image is released by the OpenMOSS Team under the [Apache 2.0 License]. These files are a quantized derivative and are distributed under the same license — see the original repositories for full terms, permitted uses, and any commercial-use restrictions.