MOSS-Image went from a fresh release on October 16, 2026 to one of the busiest open image models in ComfyUI in under a month. It now has fine-tunes, 4-step and 8-step speed LoRAs, an edit LoRA, an fp16 fix, three LoRA trainers and dozens of community workflows. That is great news, but it also makes it hard to know where to start. I have written a separate deep guide for each part. This page puts them in order, gives you the short answer for each step, and links to the full walkthrough when you need it.
- What: a 5B open text-to-image model from OpenMOSS (Apache 2.0) that uses Qwen3.5-9B as its text encoder and the FLUX.2 VAE. It renders text and follows JSON bounding-box layouts.
- Files for 12-16 GB:
moss_image_sft_int8_convrot.safetensors(5.4 GB) indiffusion_models,qwen3.5_9b_int8_convrot.safetensors(9.75 GB) orqwen3.5_9b_w4a8_gptq.safetensors(7.3 GB) intext_encoders,flux2-vae.safetensorsinvae. - Settings: Euler, 50 steps (25 for drafts), CFG 3, shift 5, skip layer guidance on layer 22, 1-2 megapixels in multiples of 16.
- Faster: 4-step rapid-moss or 8-step DMD LoRA on the bf16 SFT model with CFG 1, or the SPEED sampler (about 1.3-1.5x) with no new weights.
- Edit photos: MOSS-Image Edit LoRA with up to 4 reference images.
- Black images in fp16: switch to MOSS-Image-heal.
- Train your own: KohakuMoss (6 GB+), Fizgig 9.2 (about 20 minutes on an RTX 5090) or ZenTrainer inside ComfyUI.
What Is MOSS-Image?
MOSS-Image is an open-weights image model family from OpenMOSS, the research group behind the MOSS language models (Fudan University, the Shanghai Innovation Institute and MOSI Intelligence). It is a 5B single-stream diffusion transformer trained from scratch on about 3.2 billion images. Two checkpoints shipped on day one: MOSS-Image-Base for training and MOSS-Image-SFT for everyday generation.
Three things make it different from FLUX or Qwen-Image:
- A huge text encoder. Prompts are read by a full Qwen3.5-9B language model, almost twice the size of the image model. That is why it handles long prompts (up to about 4K tokens), English and Chinese text, and booru-style tags.
- JSON layouts. You can tell it where each object, character and line of text goes with bounding boxes.
- A fast community. ComfyUI added native support within a day (PR #17095, by Kijai), and speed LoRAs, trainers and fine-tunes followed within a week.
It shares the FLUX.2 VAE with FLUX.2 and FLUX 3 Image Klein, so if you already run those, one file is already on your drive.
The MOSS-Image Roadmap: Which Guide Do You Need?
| I want to… | Read this guide | Short answer |
|---|---|---|
| Set up MOSS-Image in ComfyUI | MOSS-Image in ComfyUI: Files, Settings, Layout JSON and 4-Step LoRAs | Update ComfyUI, download 3 files, Euler 50 steps CFG 3 |
| Use a sharper fine-tune with an 8-step LoRA | Best Kromoss ComfyUI Workflow for 16 GB GPUs | Kromoss INT8 + DMD2 LoRA, 8 steps, CFG 1 |
| Edit photos with text instructions | MOSS-Image Edit LoRA in ComfyUI | 297 MB LoRA, up to 4 reference images, 30 steps CFG 4 |
| Fix black images or run on an RTX 20 / GTX 16 | MOSS-Image-heal: Fix Black Images in fp16 | Swap in the heal file, start with --fp16-unet |
| Render faster without new weights | ComfyUI SPEED Sampler | Manual mode, 0.5x stage for half the steps: about 1.5x |
| Train a LoRA on a small GPU | Train a MOSS-Image LoRA With KohakuMoss | 6 GB+ for adapters, export in ComfyUI format |
| Train a LoRA as fast as possible | Fizgig 9.2: Train a MOSS-Image LoRA in 20 Minutes | Batch 3 dropped from 21.7 GB to 7.1 GB in 9.2 |
| Train a LoRA inside ComfyUI | ZenTrainer Guide | Train with live samples without leaving ComfyUI |
Step 1: Pick a Checkpoint
Start with the official SFT model. Only move to something else when you have a reason.
| Checkpoint | Who made it | Pick it when |
|---|---|---|
| MOSS-Image-SFT | OpenMOSS | You want pictures. The default for generation. |
| MOSS-Image-Base | OpenMOSS | You train LoRAs or fine-tunes. Train on Base, generate on SFT. |
| Kromoss | lodestones (the Chroma author) | You want a 1024 px full fine-tune (704k images) with bundled DMD and DMD2 8-step LoRAs. Kromoss guide |
| MOSS-Image-heal | KohakuBlueleaf | You need fp16 (RTX 20 / GTX 16) or got black images. heal guide |
| Community merges on Civitai | Various | You want a ready-made look, such as photorealism or anime. Popular ones include Moody Moss Mix (25 steps, CFG 3), Realism By Stable Yogi Moss (10 steps, CFG 1, er_sde) and Jib Mix Moss Realistic. Check each card for its settings and licence. |
Tip: every checkpoint above uses the same Qwen3.5-9B text encoder and FLUX.2 VAE, so switching is just a different file in Load Diffusion Model.
Step 2: Download the Files (Exact Names and Folders)
Everything for the native workflow lives in Comfy-Org/MOSS-Image on Hugging Face. You need one model, one text encoder and one VAE.
| File | Size | ComfyUI folder | Load with |
|---|---|---|---|
moss_image_sft_int8_convrot.safetensors |
5.41 GB | models/diffusion_models/ |
Load Diffusion Model |
moss_image_sft_bf16.safetensors |
10.16 GB | models/diffusion_models/ |
Load Diffusion Model (needed for 4/8-step adapters) |
qwen3.5_9b_int8_convrot.safetensors |
9.75 GB | models/text_encoders/ |
CLIPLoader, type moss_image |
qwen3.5_9b_w4a8_gptq.safetensors |
7.34 GB | models/text_encoders/ |
CLIPLoader, type moss_image (12 GB cards) |
flux2-vae.safetensors |
336 MB | models/vae/ |
Load VAE |
Civitai names: mossImage_sft_txt_int8.safetensors and the rest
Civitai’s mirror renames the same files. The txt in a name means it is the text encoder, not the image model.
| Civitai name | Same as | Folder |
|---|---|---|
mossImage_sft_txt_int8.safetensors |
qwen3.5_9b_int8_convrot.safetensors (same SHA-256) |
models/text_encoders/ |
mossImage_sft_txt_int4.safetensors |
qwen3.5_9b_w4a8_gptq.safetensors (same SHA-256) |
models/text_encoders/ |
mossImage_sft_pruned_int8.safetensors |
SFT model, int8 | models/diffusion_models/ |
mossImage_sft_bf16.safetensors |
SFT model, bf16 | models/diffusion_models/ |
If you already have one name, you do not need the other. The full file list, including the control and edit LoRAs, is in my MOSS-Image ComfyUI guide.
Step 3: Match Your VRAM
There is no official VRAM figure. ComfyUI encodes the prompt, then offloads the text encoder before the image model runs, so what matters is that each part fits on its own.
| Your GPU | Model file | Text encoder | Notes |
|---|---|---|---|
| 8 GB | GGUF (unsloth UD-Q4_K_XL, 3.42 GB) or Kromoss Q4_K_M (3.35 GiB) | W4A8 (7.3 GB) | Use --lowvram, 32 GB system RAM, about 1 MP. Community 8 GB workflows report about 62 s for 1024×1536. |
| 12 GB | SFT INT8 (5.4 GB) | W4A8 (7.3 GB) | The leanest supported combination |
| 16 GB | SFT INT8 (5.4 GB) | INT8 (9.75 GB) | My default on an RTX 5080. bf16 + 8-step LoRA also fits. |
| 24 GB+ | SFT bf16 (10.2 GB) | INT8 or bf16 (19.3 GB) | Full precision, room for edit LoRA work (about 20 GB) |
On 16 GB, users in the Kromoss threads found INT8 plus the W4A8 encoder better than GGUF: GGUF was slower and Q8_0 is barely smaller. Keep GGUF for 8 GB cards. Apple Silicon users can try mlx-community/MOSS-Image-SFT-bf16. If you still run out of memory, my CUDA out of memory fix covers the rest.
Step 4: Settings That Matter
| Setting | Standard (50-step) | Few-step LoRA (4/8-step) |
|---|---|---|
| Sampler | Euler | Euler (scheduler simple) |
| Steps | 50 for finals, 25 for drafts | 4 or 8 |
| CFG | 3 (3-6; keep 3 for long prompts) | 1, no negative prompt |
| Shift (ModelSamplingSD3) | 5 | Remove the node for rapid-moss |
| Skip layer guidance | Layer 22, end percent about 0.5 | Off |
| Resolution | 1-2 MP, both sides multiples of 16 | Same |
Good sizes: 1024×1024, 1152×896, 1344×768, 768×1344 and 1456×1456. For text inside the image, put the exact words in quotes and keep them short. Booru-style tags work well for anime: quality tags first, then characters, then general tags. OpenMOSS also publishes an official prompt enhancer skill if you want an LLM to expand short prompts.
Step 5: Make It Faster
Fifty steps with CFG means 100 model passes per image. You have four ways to cut that:
| Method | Speed-up | Catch | Guide |
|---|---|---|---|
| rapid-moss 4-step adapter (darask0) | 16.2 s to 0.81 s DiT time on an H100 | Needs bf16 SFT and its own node; quality below 50-step | setup |
| 8-step DMD LoRA (lodestones) | 16.2 s to 1.54 s | bf16 only; lora_latest at 0.7, lora.safetensors at 1.0 |
setup |
| Kromoss + DMD2 LoRA | 8 steps, CFG 1 | Works with the INT8 Kromoss file | Kromoss guide |
| SPEED sampler | 22 s to 14-18 s (author’s test) | Calibrate with Sigma Harvest; noncommercial licence | SPEED guide |
Two newer community options are worth watching: akhaliq/MOSS-Image-IDM-2-Steps for 2-step drafts and p1atdev/MOSS-Image-SFT-KV-Cache-LoRA, which caches the text encoder projection and helps most with very long prompts (8.03 s to 5.26 s at 2,941 tokens, but slightly slower on short prompts). Both are experimental. SPEED stacks with the others, and so does SageAttention.
Step 6: Edit Photos
MOSS-Image is text-to-image out of the box. The MOSS-Image Edit LoRA by chinoll (October 22) adds instruction editing: “make it winter”, “change the jacket to red leather”, or “dress the person in image 1 in the outfit from image 2”, with up to 4 reference images. It is best at colours, outfits, lighting, seasons and style, and weaker at adding new objects.
- Native route: update ComfyUI, load
moss_image_edit_lora_rank_64_bf16.safetensorsfrommodels/loras/with a normal LoRA loader, and use the Text Encode MOSS-Image Edit node. - Node pack route: Chenkin-x/ComfyUI-Moss-Image-Edit with two ready workflows.
- Start settings: 30 steps, CFG 4, shift 5, 512×512, then go bigger.
A higher-fidelity alternative, akhaliq/MOSS-Image-Edit-HQ (armV5, October 25), raises the reference budget from 144 to 256 tokens and uses masked loss so unedited areas stay closer to the source. It wants about 20 GB in bf16, uses edit_format ai-toolkit, and needs a ComfyUI nightly build. Full steps for both are in my MOSS-Image Edit LoRA guide.
Step 7: Layout and Control
This is the feature that sets MOSS-Image apart. The core MOSS-Image Layout Prompt node takes your prompt plus boxes from Create Bounding Boxes and builds the JSON layout for you. Boxes use yxyx order in pixel offsets from the centre of the canvas.
- Type the exact output width and height into Create Bounding Boxes. Do not wire a Resolution Selector into it; that shifts the boxes.
- For colour blocks and depth maps, use the
moss_image_color_depth_control.safetensorsLoRA (313 MB,models/loras/) with the core MOSS-Image Color/Depth Control node. - Treat boxes as strong hints, not pixel-exact placement. OpenMOSS labels layout as experimental.
The JSON format with a worked poster example is in the layout section of my MOSS-Image guide.
Step 8: Train Your Own LoRA
Train on Base, then use the LoRA on SFT (or Kromoss). Three trainers support MOSS-Image well:
| Trainer | Runs as | VRAM | Best for | Guide |
|---|---|---|---|---|
| KohakuMoss | Standalone (PyPI kohakumoss) |
6 GB adapters (with caches), 24 GB full fine-tunes | Small GPUs, every LyCORIS type (LoRA, LoKr, DoRA, OFT) | KohakuMoss guide |
| Fizgig 9.2 | Standalone app, browser or RunPod | About 7.1 GB at batch 3 | Speed: about 20 min per LoRA on an RTX 5090 | Fizgig guide |
| ZenTrainer | ComfyUI custom nodes | Depends on settings | Training without leaving ComfyUI, with live samples | ZenTrainer guide |
All three export LoRAs that load with ComfyUI’s normal LoRA loader. If you train in fp16, train on MOSS-Image-heal to avoid overflow.
Common Problems and Fixes
| Problem | Fix |
|---|---|
“Unknown model type” or no moss_image option in CLIPLoader |
Update ComfyUI to the latest build (support landed in PR #17095) and restart |
| Black or broken images in fp16 | Use MOSS-Image-heal, or stay in bf16 / int8 |
| Width or height error | Both sides must be multiples of 16 (1008, not 1000) |
| 4-step or 8-step output is mushy | Load the bf16 SFT model, CFG 1, no shift node, no skip layer guidance |
| Grey noise with Kromoss | FLUX.2 VAE, EmptyFlux2LatentImage, and CFG 1 with 8 steps when the LoRA is on |
'NoneType' object has no attribute 'dtype' |
Use the core Load Diffusion Model node, not a third-party MOSS loader |
| Vertical colour stripes in flat areas | Try MATLOWAI/MOSS-Image-SFT-DeBand or the small nynxz destripe LoRA |
| Layout boxes in the wrong place | Remove the Resolution Selector and type width/height by hand |
| Out of memory on 12 GB | W4A8 encoder, about 1 MP, --lowvram |
| Garbled text | Shorter text in quotes, 50 steps; few-step LoRAs are weaker at text |
Frequently Asked Questions
Is MOSS-Image free for commercial use?
The code and the Comfy-Org repack are tagged Apache 2.0. Check the original OpenMOSS model card, and the licence of any LoRA or tool you add (the SPEED sampler, for example, is noncommercial).
How much VRAM does MOSS-Image need?
About 12 GB with the INT8 model and W4A8 encoder, 16 GB for a comfortable setup, and 8 GB with GGUF and --lowvram. Plan for 32 GB of system RAM.
What is mossImage_sft_txt_int8.safetensors?
The Civitai name for the Qwen3.5-9B int8 text encoder, the same file as qwen3.5_9b_int8_convrot.safetensors. It goes in models/text_encoders/.
Should I use Base or SFT?
SFT for generating images, Base for training LoRAs.
Is Kromoss better than MOSS-Image-SFT?
It is sharper at 1024 px and ships with an 8-step LoRA, but community feedback is mixed. Try both on the same seed.
Can MOSS-Image edit existing photos?
Yes, with the MOSS-Image Edit LoRA, using up to four reference images.
Does MOSS-Image work on a Mac?
Yes, through the MLX conversion or GGUF builds, but it is much slower than on an NVIDIA GPU.
Final Thoughts
Start simple: SFT INT8, one text encoder, the FLUX.2 VAE and the standard 50-step settings. Once that works, add the 8-step LoRA for speed, the Edit LoRA for photo edits and KohakuMoss or Fizgig when you want your own style. MOSS-Image is moving fast, so I update this page whenever I add a MOSS-Image guide. Bookmark it, and if you also make video, my MiniMax H3 ComfyUI guide is the matching hub for that side.
