MOSS-Image Review: Text, Layout Control, NSFW, VRAM and ComfyUI

MOSS-Image is an unusually useful open image model for posters, illustrated information cards and scenes that need explicit composition. Its strongest proposition is the combination of long prompts, text rendering and native JSON layout control—not a guarantee that every anime character or style will look better than a specialist model.

Released on October 16, 2026, the Base and SFT checkpoints are available to download. Start with SFT for text-heavy design experiments; keep Base in consideration for stylistic exploration and fine-tuning. The main costs are a substantial text encoder, a standard 50-step workflow and the time needed to check typography and details. Direct reference-image editing is not part of the released backbone.

Verdict and key specifications

Developer / release	OpenMOSS · October 16, 2026
Model components	5B single-stream DiT + Qwen3.5-9B encoder + FLUX.2 small-decoder VAE
Input	Natural language, tag sequences, JSON layout · up to 4,096 tokens
Suggested starting output	1024 × 1024; recommended 1–2 MP; dimensions divisible by 16
Baseline sampling	Euler · 50 steps · CFG 3 · shift 5 · uncond skip layer 22
Weights / access	Apache-2.0 weights · local inference and official online demo

Reasons to try it

• Explicit placement for text and objects instead of relying entirely on “left” and “right” in prose.
• A broad visual range covering photography, illustration and anime, with detailed language and tag inputs.
• Downloadable Base/SFT weights and an emerging quantization and LoRA ecosystem.

Reasons to keep alternatives

• The 5B headline excludes the large text encoder; low-memory deployment needs careful choices.
• Short prompts, named characters, artist tags, fingers and exact logical constraints can remain inconsistent.
• No native image-conditioned editing; few-step community adapters are separate, changing workflows.

Base vs SFT: which checkpoint should you download?

SFT is fine-tuned from Base on curated data. OpenMOSS’s whiteboard examples show clearer text grouping and organization in SFT, making it a sensible first checkpoint for signs, posters and information graphics. This is a practical starting recommendation, not proof that SFT wins every prompt. SFT model card

Base preserves an unrefined foundation for adaptation. The official guide recommends it for community fine-tuning, and a developer in the AMA personally preferred its more realistic look. “More free” should be understood as a different output distribution, not a formal promise of unrestricted content or superior quality. Base model card

The two official examples below are developer-selected illustrations of the versions. Their composition and subject appearance differ, so they should not be read as a controlled identity-consistency test.

Image quality, anime knowledge and multilingual text

For design work, text and placement are more persuasive than a vague claim of “better aesthetics.” The technical report gives MOSS-Image a 96.3 Mean-6 text-coverage score on LongLongTextBench, a 240-task English/Chinese benchmark. This is developer-reported coverage, not a 96.3% guarantee of correct spelling, attractive typography or Japanese accuracy. Its RealLayoutArena system ranks second in the report’s end-to-end comparison; a separate shared-enhanced-prompt test ranks its backbone first. Those protocols are not interchangeable. Technical report

Japanese signage and dialogue have encouraging user reports, including a 35-prompt community test. They do not establish reliable manga lettering across fonts, long sentences or vertical writing. English is the official guide’s preferred prompt language; for a critical title, quote the exact text, insert deliberate line breaks and inspect every character. For Spanish, Portuguese, French and German, test accents and punctuation explicitly rather than extrapolating from the Chinese benchmark. Community test and settings

Anime is supported, but named-character knowledge and style control are different questions. Tag support does not establish complete Danbooru coverage. Users in the AMA report weak artist-name response; a Japanese hands-on article found anime consistency sensitive to style wording. Backgrounds and separated subjects can be promising without making MOSS-Image a universal replacement for Anima. Hands, small faces, negation and repeated attributes still deserve close inspection. Developer AMA and user feedback Japanese hands-on examples

A polished or “GPT-like” appearance is a style observation, not evidence of synthetic training data. The report describes deduplication and an internal detector reducing roughly five billion source images to about three billion. That is the team’s account of corpus preparation, not an independent audit or a guarantee of complete dataset clearance. Training-data description

JSON and bounding boxes: the feature that matters most

Layout is the most concrete reason to add MOSS-Image to a local workflow. A structured prompt can describe the overall scene, then give individual text or object items their own captions and boxes. This is useful for a heading above a product, two separated characters or a panel-based composition. JSON is optional; natural language still works. Official input guide

The coordinate convention is easy to get wrong. With abs_centralized and yxyx, a box means [y_min, x_min, y_max, x_max] in pixels relative to the canvas centre. On a 1024-square image, each axis runs from −512 to +512. Do not paste ordinary top-left coordinates or a community 0–1000 layout into this schema without conversion.

The following original input example places a title above a mountain. Set output to 1024 × 1024 and paste the complete JSON into the positive prompt. Boxes guide generation; they are not vector layers or guaranteed pixel-perfect placement. Changing JSON and regenerating does not preserve an uploaded photograph. The technical report explicitly lists direct image conditioning and image editing as future work.

{
  "summary": {
    "meta": [
      "poster"
    ],
    "canvas": "single",
    "caption": "A minimalist travel poster with a cream background, a blue mountain illustration below, and the exact title SUMMER MARKET at the top."
  },
  "items": [
    {
      "label": "text",
      "caption": "Large dark-blue title centered at the top.",
      "text": "SUMMER MARKET",
      "bbox": [
        -416,
        -384,
        -288,
        384
      ]
    },
    {
      "label": "object",
      "caption": "A flat blue mountain illustration in the lower center.",
      "bbox": [
        -192,
        -288,
        384,
        288
      ]
    }
  ],
  "bbox_coord_space": "abs_centralized",
  "bbox_format": "yxyx"
}

VRAM, download sizes and generation speed

Download size and peak VRAM are different measurements. In the Comfy-Org file listing checked on October 20, the BF16 backbone is about 10.16 GB, the BF16 text encoder 19.31 GB and the small VAE 0.25 GB: approximately 29.72 GB of weight files before the environment and cache. Offloading can avoid keeping them all on the GPU simultaneously. Current component files

For 8–12 GB cards, start with a compatible quantized workflow and CPU offloading, at roughly one megapixel and one image per batch. More GPU memory gives room for larger workloads, but “16 GB comfortable” is not a universal specification and full-resident BF16 can remain tight even on a 32 GB card once working memory is added. System RAM, prompt length and VAE decoding matter too.

One RTX 3060 12 GB user reported about 9.7 GB peak VRAM and 137 seconds at one megapixel for a 50-step run, versus about 24 seconds with an early 8-step DMD setup. These are attributed results from a particular earlier third-party quant and custom workflow, not current Comfy-Org performance promises. The quant creator later recommended moving to Comfy-Org files. Do not transfer those times to another GPU, resolution or adapter. Hardware report and follow-up

Approximate decimal GB on disk, not minimum VRAM. W4A8 here describes the encoder, not the whole pipeline.

How to start in ComfyUI without mixing workflows

Use one complete route. For current native ComfyUI support, update ComfyUI, the frontend and templates, search templates for MOSS-Image, and use the matching Base or SFT workflow with Comfy-Org component files. Native support was merged on October 16; older project text saying integration is still in progress does not reflect that merged PR. ComfyUI integration

The OpenMOSS plugin is another route: its HF Loader expects a local Diffusers directory containing transformer/, text_encoder/, tokenizer/ and vae/. A Hub repository name or a lone backbone file is not that directory. Do not combine a plugin graph with an unrelated native loader and assume their sampling patches are equivalent. Official plugin instructions

Start at 1024 × 1024, 50 steps, CFG 3, shift 5, Euler / sgm_uniform and negative-branch layer skipping at 22 when the chosen workflow exposes those controls. Keep standard SLG off unless deliberately replacing that guidance mode. Check the VAE: OpenMOSS recommends its small decoder to reduce certain artifacts. If you see noise, spots or missing-node errors, restore the matching official template and components before stacking speed patches.

Example file locations for the native single-file route; choose one matching backbone and encoder, not every precision at once.

ComfyUI/models/
  diffusion_models/
    moss_image_sft_bf16.safetensors
    # OR moss_image_sft_int8_convrot.safetensors
  text_encoders/
    qwen3.5_9b_bf16.safetensors
    # OR qwen3.5_9b_int8_convrot.safetensors
  vae/
    flux2-vae_small.safetensors

Prompts that give MOSS-Image a clearer brief

Give the model a coherent design brief: subject, action, perspective, spatial relations, lighting, style and exact visible words. A detailed brief is not a requirement to fill all 4,096 tokens. Contradictory instructions and lists of unrelated quality terms create more ambiguity.

For tags, the guide recommends Quality → Score → Meta → Year → Character → Series → Artist → General. Quality terms can affect the result, but “8K” in a prompt does not change output dimensions and an artist token is not a guaranteed style switch. Pair important character names with visible descriptions. Prompt recommendations

For English artwork, specify the title as an exact quoted string and choose intentional line breaks. For Japanese artwork, a direct Japanese scene description is worth testing; an English scene brief with the exact Japanese lettering quoted separately is a useful fallback. Keep a fixed seed while adjusting the brief, then test several seeds before judging consistency.

Original, safe prompt to adapt

A clean editorial illustration of two adult market vendors behind separate stalls. The vendor on the left wears a red apron and holds one yellow lemon; the vendor on the right wears a blue apron and holds one green pear. Warm morning light, restrained colours, a clear space above them containing exactly "SUMMER MARKET". No additional words.

Quantization, GGUF, LoRAs and acceleration

INT8 ConvRot files are available from Comfy-Org, and community projects provide lower-bit and GGUF variants. A GGUF backbone, a GPTQ encoder and a BF16 model are not interchangeable file formats. Check the loader, encoder and VAE specified by the quantization author, then compare small text and fine details against your baseline. Community GGUF model card

The 8-step DMD release is a community work in progress, not an official MOSS Turbo checkpoint. Its current card gives LoRA-loader instructions and marks an older custom data-space / FP32 section deprecated. Follow the current matching adapter and workflow instead of copying that older section or merely reducing a standard sampler to eight steps. Spectrum and four-step experiments should also be assessed as separate speed/quality trade-offs. Current DMD instructions

For training, the official guide recommends Base and freezing the text encoder. Caption the visible content accurately and preserve exact image text; do not assume Anima, SDXL or FLUX LoRAs can be reused because their filenames share “LoRA.” The ecosystem is moving quickly, so model compatibility, training examples and sample settings matter more than the number of downloads.

NSFW, commercial use and the real cost

Local users report adult/NSFW generation from both the family and its derivatives, but “uncensored” is not a reliable description of every hosted endpoint. A hosting service or external prompt enhancer can apply its own rules. Capability reports do not guarantee correct anatomy, a content policy or permission to use a person’s likeness. For SFW work, the guide suggests an NSFW negative tag among other quality negatives; still inspect the output. Community discussion Official negative-prompt guidance

The released weights are labeled Apache-2.0, and the standalone pipeline code uses Apache-2.0 as well. This supports commercial model use subject to the license conditions; third-party components and downloaded adapters retain their own terms. It does not establish that every generated character, logo or likeness is cleared for a commercial campaign. Model license

There is no single official per-image price for running local weights. Budget for GPU time, storage, electricity and rejected takes; paid hosts set their own rates. The official demo is a way to explore availability, not a promise of permanent free capacity. Start from the OpenMOSS project or its linked Hub repositories rather than treating a similarly named generator site as the developer’s service.

MOSS-Image vs Anima, Krea 2, Qwen and FLUX

Choose by task, not by launch-week slogans. A community comparison used the same prompts but different hardware and step counts—MOSS on an H100 at 50 steps, Krea Turbo on a Mac at four steps—so it cannot establish a speed winner. There is also no evidence here that Anima development has stopped. Comparison setup

The anime version: a plan, not a released upgrade

In the developer AMA, the reply to a Turbo question says the current priority is an anime version; another reply says anime was only a small fraction of the existing training mix. This makes specialist anime tuning a credible development direction, but it supplies neither a release date nor downloadable anime-version weights. Read the AMA

Do not buy hardware or retire an established workflow on the assumption that this update, a native editor or an official Turbo version is already available. Evaluate the released checkpoint, and treat future releases as new products that need their own tests.

Practical questions before downloading

Is MOSS-Image free and commercially usable?

The downloadable weights are Apache-2.0; local operation still has hardware and running costs. Hosted pricing is separate, and components or LoRAs can have additional terms. Keep the license and exact model revision with a commercial project.

Should beginners choose Base or SFT?

Start with SFT for typography and organized design, then try Base when its style fits better or when training an adapter. Use the same prompt and several seeds; one attractive example is not a reliable version comparison.

Can it run on a 12 GB GPU?

A documented user setup did, using quantization and offloading. That does not mean full BF16 fits or all resolutions are practical. Match the workflow, monitor VRAM and system RAM, and start with one image at around one megapixel.

Is Japanese lettering guaranteed?

No. Community results are encouraging, while the official long-text benchmark covers English and Chinese. Test your actual glyphs, line breaks and speech bubbles, and proofread the exported image. Use a layout editor for text that must be exact.

Can I edit a reference photo or keep a character identical?

The released backbone does not directly condition on reference images. Reusing a seed and changing JSON can guide regeneration, but it is not a guarantee of input preservation or character identity. Choose an image-editing model for that requirement.

Why does an eight-step setup look worse?

The adapter may not match the checkpoint, precision or sampling workflow. Establish a standard 50-step baseline first, then follow the adapter’s current card. Fast generation can lose lighting nuance, fine constraints or text quality.

Who should use MOSS-Image now?

MOSS-Image is worth adding if you want local, text-rich image generation with explicit composition and a broadly permissive model license. SFT is a useful starting point for posters and information cards; Base remains valuable for adaptation and a different visual distribution. Neither version earns a blanket “Anima killer” label.

Before adopting it, run a small test pack: a title with punctuation, two differently dressed characters, a fixed layout and a fine-detail scene. Record checkpoint, encoder, VAE, precision, prompt, seed and time including encoding. Choose the workflow that produces usable results consistently—not the most flattering launch screenshot.

Evidence and evaluation scope

This is a launch review of released files, official examples, documented workflows and attributed community testing, checked through October 20, 2026. It is not an AirMore GPU benchmark. Published developer samples are labeled, and community timings are not presented as our measurements.

Useful primary sources and reproducible reports

• Project and downloads
• Official user guide
• Technical report and benchmark protocol
• Comfy-Org files
• Developer AMA
• Community hardware test
• Japanese hands-on report
• Community acceleration card