TRY ONLINE FREE

Qwen Image Editor: 2 Free Generations, No VRAM Needed!

Home/Blog/MOSS Image Guide & Workflows
Open Source AI & Visual Workflows

MOSS Image Guide: ComfyUI Workflow, Prompting & Edit

Master the OpenMOSS 5B Diffusion Transformer: ComfyUI low-VRAM setups, bilingual typography prompts, Base vs SFT selection, and in-browser localized image editing.

Elena Rostova
Elena RostovaVerified Specialist
Lead Visual AI & Generative Media Specialist
2026-10-2110 min read
Bottom Line Up Front (BLUF) Overview

MOSS Image is an open-weights, 5-billion parameter Diffusion Transformer (DiT) model developed by the OpenMOSS Team (Fudan University & Shanghai Innovation Institute) under Apache 2.0. Operating on an advanced multilingual text encoder, MOSS Image excels at English and Chinese typography rendering and structured JSON layout control. While local ComfyUI inference requires 10GB–24GB VRAM, creators seeking zero-install visual inpainting, conversational text editing, and instant refinement can utilize Qwen Image Editor directly in any modern browser without local setup.

What is MOSS Image? Base vs SFT Models Explained

While models like FLUX.1 and Midjourney v6 generate striking visuals, rendering legible bilingual typography has remained a challenge. MOSS Image tackles this bottleneck through a unified 5B Diffusion Transformer backbone with deep multilingual character modeling.

When accessing the model on HuggingFace (OpenMOSS-Team/MOSS-Image), creators encounter two primary checkpoints:

MOSS-Image Base

Pre-trained on broad-scale visual corpora. Preserves artistic style entropy and raw conceptual richness. Ideal for researchers training LoRAs or building custom fine-tunes on Civitai.

Best For: LoRA training, style distillation, custom architectural backbones.
MOSS-Image SFT

Supervised Fine-Tuned on curated instructional datasets. Drastically improves prompt adherence, spatial reasoning, bilingual text clarity, and native JSON coordinate layout alignment.

Best For: Production poster design, ecommerce typography, direct ComfyUI workflows.
Interactive MOSS Image Prompt & Bounding-Box Generator

Configure bilingual typography copy, preview coordinate bounding boxes, and generate ComfyUI-ready MOSS Image JSON inputs.

Live In-Browser Tool
OpenMOSS Layout Payload (JSON)
{
  "prompt": "Commercial promotional poster. Bold typography reading \"FUTURE OF VISION 2026\" and Chinese bilingual typography \"下一代智能多模态视觉\". Clean graphic layout, volumetric cinematic lighting, photorealistic 8k render.",
  "model_target": "OpenMOSS/MOSS-Image-SFT",
  "layout_control": {
    "enabled": true,
    "regions": [
      {
        "content_type": "bilingual_text",
        "text": "FUTURE OF VISION 2026 / 下一代智能多模态视觉",
        "bounding_box_normalized": [0.08, 0.10, 0.28, 0.90]
      }
    ]
  },
  "inference_settings": {
    "steps": 10,
    "guidance_scale": 4.5,
    "quantization": "INT8 ConvRot / GGUF Q4_K_M"
  }
}
Hardware Advice: Must use official MOSS Image compact Decoder to prevent color spotting

MOSS Image Prompting Guide: Bilingual Text & Layout Control

Generating clean typography in MOSS Image requires structuring prompts properly to prevent letter distortion. Drawing from community testing, here are the core prompting principles:

1Explicit Quotation Framing for Text Entities

Encapsulate target text in straight quotation marks ("...") preceded by explicit design directives such as typography reading "...", signage with the word "...", or billboard lettering "...".

2Chinese-English Dual Hierarchy Styling

When designing bilingual posters, specify stylistic hierarchies. Define whether the Latin title serves as the primary hook and Chinese text serves as the descriptive seal, or vice versa.

Prompt Template: "Cosmetics tube, primary bold English font reading "PURE BOTANICS", centered Chinese subtitle calligraphy reading "植萃精华", matte paper texture, studio softbox lighting."
3Bounding-Box Coordinate Formulation (SFT Model Only)

MOSS-Image SFT natively recognizes normalized coordinates [ymin, xmin, ymax, xmax] between 0.0 and 1.0. Defining a bounding box anchors the generative latent distribution, preventing text from floating unpredictably across visual subjects.

MOSS Image ComfyUI Setup, Workflows & Low-VRAM Optimization

For local creators, running inside ComfyUI provides granular node control. However, community discussions on Reddit and GitHub highlight critical setup details:

Critical Bug Alert: VAE Spotting & Decoding Artifacts

Many users attempting to run MOSS Image with standard FLUX or SDXL VAE decoders encounter color banding and magenta artifacts. The architecture relies on a specialized compact decoder. Ensure you load the official moss-compact-vae.safetensors node inside your ComfyUI loader chain.

Recommended ComfyUI Pipeline Node Script
# ComfyUI OpenMOSS Custom Pipeline Node
from comfy_moss import MossImageLoader, MossLayoutControl, CompactVaeDecoder

# Load official OpenMOSS weights (Base or SFT)
model = MossImageLoader.load_checkpoint(
    variant="MOSS-Image-SFT",
    precision="int8_convrot"
)

# Connect native compact MOSS VAE decoder to prevent spotting
vae = CompactVaeDecoder.load_decoder("moss-compact-vae.safetensors")

# Inject structured bounding box coordinates
layout = MossLayoutControl.create_region(
    bbox=[0.08, 0.10, 0.28, 0.90],
    text_content="FUTURE OF VISION 2026"
)
How to Run on 10GB–12GB VRAM (INT8 & GGUF)
  • INT8 ConvRot Quantization: Squeezes model weights to ~9.7GB VRAM for smooth RTX 3060 (12GB) execution.
  • DMD 8-Step Turbo Acceleration: Drops inference latency from 28 steps (35s) to under 8 seconds per generation.
  • Offloading Text Encoder: Offload text encoder to system CPU RAM after prompt tokenization to free up 2.5GB VRAM for diffusion attention passes.

Why MOSS Image Needs an Online Editor for Inpainting & Text Edits

While The 5B model represents a breakthrough in text-to-image synthesis, production work requires post-generation refinement. Creators face three common hurdles:

Typo in an Otherwise Perfect Image

If the model generates a stunning visual but misplaces a letter, re-running the prompt burns GPU compute and alters the entire image composition.

Lack of Native Conversational Inpainting

It is strictly a forward text-to-image generator. It does not provide localized multi-turn conversation masks to swap a model's outfit or extend canvas aspect ratios.

Hardware & Setup Fatigue

Creators often prefer to avoid Python environments, Git pulls, CUDA out-of-memory crashes, or expensive cloud instance fees.

The Ideal Solution: Pair MOSS Image with Online Editing
Edit, Inpaint & Refine MOSS Image Creations in Your Browser

Take your generations and refine them with zero VRAM hardware requirements. Use conversational inpainting to replace text, modify visual elements, or isolate product backgrounds with our AI Background Remover.

Open Online Studio Free

MOSS Image Benchmark Matrix: MOSS Image vs Qwen Image vs FLUX.1

To assist visual creators in selecting the right foundation system, we have compiled an objective benchmark matrix:

Evaluation DimensionMOSS Image (5B DiT)Qwen Image (Online)FLUX.1 (Dev/Schnell)
License & Code AccessApache 2.0 (Open Weights)Open Foundation / Cloud StudioNon-commercial (Dev) / Apache 2 (Schnell)
English TypographyExcellent (Latin signage & posters)Industry-Leading (Crisp rendering)High (Latin text only)
Chinese (Hanzi) TypographyOutstanding (Precise strokes)World-Class (Native VLM encoder)Poor / Unsupported natively
Bounding-Box Layout ControlNative JSON Coordinates (SFT)Conversational Visual PromptsRequires external ControlNet / LoRA
Localized Image InpaintingComplex ComfyUI node wiringNative In-Browser Click & EditSupported via FLUX-Fill models
VRAM Requirement (Local)10GB (INT8) to 24GB (BF16)0GB VRAM (100% Cloud In-Browser)12GB (NF4) to 24GB (FP16)

MOSS Image Frequently Asked Questions (FAQ)

What is MOSS Image and is it open source on HuggingFace?

MOSS Image is a 5-billion parameter open-weights Diffusion Transformer (DiT) text-to-image foundation model developed by the OpenMOSS Team (Fudan University and Shanghai Innovation Institute). Released under the Apache 2.0 license, both Base and SFT checkpoints are available on HuggingFace and ModelScope.

What is the difference between MOSS-Image Base and MOSS-Image SFT?

MOSS-Image Base is the foundation model trained on broad visual-text corpora, making it ideal for downstream research, fine-tuning, and creating custom community LoRAs. MOSS-Image SFT is the Supervised Fine-Tuned instruction model optimized for prompt adherence, dual-language typography rendering, and native JSON layout control.

How to fix the VAE spotting and artifact bug in ComfyUI for MOSS Image?

A common bug occurs when ComfyUI workflows default to using FLUX.1 VAE decoders with OpenMOSS, causing color splotches and magenta artifacts. To resolve this, install the official 'comfyui/moss-image' package and connect your workflow to the specialized compact MOSS VAE decoder checkpoint.

Can I run MOSS Image with low VRAM on 10GB or 12GB consumer graphics cards?

Yes. While uncompressed BF16 requires 24GB VRAM, the community provides INT8 ConvRot and GGUF quantizations. Paired with an 8-step DMD Turbo acceleration node, peak VRAM consumption drops to approximately 9.7GB–10GB, running smoothly on consumer cards like the RTX 3060 (12GB) or RTX 4060.

How can I edit text or fix mistakes in creations online without GPU hardware?

MOSS Image specializes in text-to-image generation but lacks native interactive inpainting, multi-turn conversational editing, and localized text correction. If your generated image has a typo or needs background replacement, you can use Qwen Image Editor (qwenimage-editor.com) directly in your browser with zero local VRAM requirements.

Elena Rostova
Elena RostovaAuthor & Technical Reviewer
Lead Visual AI & Generative Media Specialist

Former VFX technical director and generative media researcher specializing in multimodal diffusion pipelines, neural typography, and visual inpainting workflows. Elena leads multimodal diffusion testing and bilingual generative typography research, developing resilient inpainting pipelines for ecommerce and enterprise visual studios.

Ready to Edit & Refine Your AI Visuals Online?

Skip local hardware barriers, VAE errors, and 24GB VRAM requirements. Upload your designs and experience seamless conversational AI image editing and bilingual typography in your browser today.