MOSS-Image SFT INT8 (Abiray) for ComfyUI

MOSS-Image is a 5B-parameter Diffusion Transformer from OpenMOSS, built for strong in-image text rendering and controllable layouts. This is a mirror of Abiray’s INT8 ConvRot quantization of the SFT checkpoint. It is the version that loads cleanly with the comfyui-moss-Image custom node.

Base vs SFT

• Base: Full training signal, recommended starting point for LoRA training.

• SFT (this file): Fine-tuned on curated data. Produces cleaner text grouping and more organised layouts. Recommended for generation.

Key strengths

• Accurate English and Chinese text inside images (up to ~4K token prompts).

• JSON layout control for precise placement of text and objects (abs_centralized + yxyx).

• Works with natural language, tags, or layout JSON.

• Photoreal and stylised output.

• INT8 keeps VRAM usage low while staying close to the BF16 quality.

Required companion files
You also need these two files:

• Text encoder: qwen3.5_9b_int8_convrot.safetensors

• VAE: flux2-vae.safetensors

Place the diffusion model in models/diffusion_models, the text encoder in models/text_encoders, and the VAE in models/vae.

Recommended settings

• Sampler: Euler

• Scheduler: sgm_uniform

• Steps: 30-50

• CFG: 3.0 (3–6 range)

• Shift: 5.0

• SLG: Uncond skip on layer 22 (24/26 also useful)

• Resolution: multiples of 16, 1–2 MP preferred

ComfyUI
Works with the custom node by wanaigc (https://github.com/wanaigc/comfyui-moss-Image) and with native support (ComfyUI ≥ 0.41.3 with the MOSS-Image nodes). Keep SLG on for complex scenes to reduce anatomy and detail collapse.

A working sample workflow is attached, and I have published my own refined workflow already.

Original models and code: OpenMOSS/MOSS-Image. Quantization by Abiray. Apache 2.0. All credit to the OpenMOSS team.