MOSS-Image-SFT-Turbo INT8 ConvRot AIO

Experimental model files for ComfyUI and others, mainly created for testing in Radiant Canvas for fast macOS inference (https://apps.apple.com/us/app/radiant-canvas-ai-image-gen/id6802973075).

This is the all-in-one (AIO) version of toxicdog/MOSS-Image-SFT-Turbo-INT8-ConvRot, a fast 8-step variant of OpenMOSS-Team/MOSS-Image-SFT. One checkpoint file contains the diffusion model, the text encoder and the VAE, so ComfyUI needs only a single Load Checkpoint node.

Use CFG 1 and 8 steps. No negative prompt is used.

File

checkpoints/moss_image_sft_turbo_int8_convrot_aio.safetensors (15.5 GB) contains:

The diffusion model is the same as in the separate-files repository.

Using it in ComfyUI

MOSS-Image is supported natively in ComfyUI master since October 16, 2026. No custom nodes are needed. Use a ComfyUI build that includes MOSS-Image support and INT8 ConvRot.

1. Place the file in:

   ComfyUI/models/checkpoints/moss_image_sft_turbo_int8_convrot_aio.safetensors
   

2. Build the workflow:

   Load Checkpoint      MODEL -> KSampler (model)
                        CLIP  -> CLIP Text Encode (your prompt) -> KSampler (positive)
                        CLIP  -> CLIP Text Encode (empty)       -> KSampler (negative)
                        VAE   -> VAE Decode (vae)
   Empty Flux 2 Latent  LATENT -> KSampler (latent_image)
   KSampler             LATENT -> VAE Decode (samples) -> Save Image
   

   • Load Checkpoint: select this file. Its MODEL, CLIP and VAE outputs replace the separate diffusion model, text encoder and VAE loaders.
   • CLIP Text Encode: your prompt for the positive input. The negative input needs a conditioning, but it is ignored at CFG 1. An empty prompt is fine.
   • Empty Flux 2 Latent: the image size. Width and height must be multiples of 16. About 1 megapixel (for example 1024 × 1024) works best.

3. Set the KSampler:

   Setting	Value
   Steps	8
   CFG	1
   Sampler	euler
   Scheduler	simple
   Denoise	1.0

The model uses the standard flow-matching sampler with the default MOSS-Image shift of 5, which ComfyUI applies automatically. No custom scheduler or ModelSampling node is needed.

Memory

The checkpoint is 15.5 GB: about 5.4 GB for the diffusion model, 9.7 GB for the text encoder and 0.3 GB for the VAE. The text encoder only runs when the prompt changes, so ComfyUI can move it out of GPU memory while sampling. On smaller GPUs, ComfyUI's automatic model offloading keeps it working, at some cost in speed.

Separate files instead

If you already have the MOSS-Image text encoder and VAE, or want to swap them, use the separate diffusion model file from toxicdog/MOSS-Image-SFT-Turbo-INT8-ConvRot with ComfyUI's diffusion model, text encoder and VAE loaders instead.

Precision

• Diffusion model: the 252 attention and MLP block weights are INT8 ConvRot (per-channel scale, ConvRot group size 256), quantized from FP32 weights. rope.bands is FP32. All other tensors are BF16.
• Text encoder: Comfy-Org's INT8 ConvRot Qwen3.5-9B, unchanged.
• VAE: Comfy-Org's FLUX.2 VAE in FP32, unchanged.

This is an experimental release. Check output quality for your own prompts before relying on it.

License

Released under the Apache License 2.0. The diffusion model is a modified version of MOSS-Image-SFT by OpenMOSS. The text encoder (Qwen3.5-9B by Qwen) and the VAE (the FLUX.2 VAE by Black Forest Labs, released under Apache 2.0 with FLUX.2 [klein] 4B) are included unchanged and are also licensed under the Apache License 2.0. A copy of the license is included in this repository.