🎨 V67 Moss Image Aesthetic — The Native Aesthetic LoRA Component for MOSS-Image

Premise: If you use my model and your image is chosen for the CIVITAI cover, you receive 500 Yellow Buzz — even in combination with other models.
R+ images are hidden from the gallery. Trained with MOSS-Image.

Putting a Like on the V67 models is worth more than giving me Buzz (yellow/green).

📌 Introduction: The House of MOSS-Image

Welcome to the next evolution of open-weight image generation. This LoRA has been designed exclusively for use with MOSS-Image, the compact 5B text-to-image model from OpenMOSS and MOSI Intelligence.

MOSS-Image was open-sourced on 16 October 2026 under the Apache 2.0 license — a genuinely permissive license with no revenue thresholds and no separate commercial tier. The technical report lists MOSI Intelligence, the Shanghai Innovation Institute (SII), Fudan University, and Shanghai Jiao Tong University. Within days of release, the model topped Hugging Face's text-to-image trending list, with the official repository climbing past roughly 1.4 thousand GitHub stars.

MOSS-Image is built around a single-stream diffusion transformer (DiT) with a frozen Qwen3.5-9B text encoder and a FLUX.2 variational autoencoder, feeding a unified generation path. The training recipe deliberately filters out AI-generated images to preserve photographic variety — a choice that gives the model a less "artificial" feel than many of its contemporaries.

Its headline features:

✍️ Dense text rendering — in English and Chinese. MOSS-Image accepts prompts up to 4,096 tokens and is engineered for long-form text rendering. On the team's LongLongTextBench (LLTB) — 240 literary-text rendering tasks (120 English, 120 Chinese) with a mean target length of roughly 1,340 characters — it reports a Mean-6 text coverage of 96.3% among open-weight models. In the authors' own runs, this compares with 64.9% for Qwen Image 2512 and 11.8% for FLUX.2.

📐 Native JSON layout control. Objects, rendered text, and their spatial positions can be specified in a structured JSON prompt with bounding boxes. This is a first-class capability, not an afterthought — it supports layout-to-image generation, text editing, and element addition/removal in a structured way. On RealLayoutArena, an end-to-end blind design test, MOSS-Image reports an Elo of 1,698.70 with short briefs (second behind GPT Image 2) and 1,807.34 with shared enhanced prompts, topping that group in the team's own scoring.

⚙️ Compact, adaptable backbone. A 5B DiT designed for efficient generation and efficient fine-tuning. This is why community LoRAs for MOSS-Image are already appearing at ~70 MB and training successfully on a single RTX 4090 in about an hour.

🖼️ Flexible resolution and aspect ratios. Recommended outputs at 1M, 1.5M, or 2M total pixels, with support from 0.25M (512×512) up to 4M (2K), in arbitrary aspect ratios aligned to multiples of 16.

🚀 Turbo and distilled variants. The community has already shipped 8-step and 4-step distillations, quantization packs (INT8, GGUF), and ComfyUI workflows, making rapid iteration practical on consumer hardware.

🏆 Caveat on benchmarks: the LLTB and RealLayoutArena figures are self-reported by the model's authors or crowd-sourced. They should be read as directional signals, not independent lab audits.

🚀 The Project: V67 Moss Image Aesthetic

This LoRA component is designed to enhance MOSS-Image's native capabilities, delivering:

Enhanced Human Realism — Natural skin texture, individual hair strands, authentic expressions. It works against the residual smoothness that can appear in base-model portraits.
Natural Texture Fidelity — Sharp landscapes, detailed fur, flowing water, dense foliage — preserving detail where diffusion models typically soften.
Text-Native Synergy — MOSS-Image already renders text better than most open models. This LoRA is tuned not to fight that strength: it keeps typography crisp and legible rather than degrading it into stylised noise. Designed with posters, menus, e-commerce images and advertising materials in mind.
Production-Ready Quality — Built for marketing assets, e-commerce, catalogues, and professional content where layout and lettering matter.
Versatility — Works across generic subjects without overfitting on a single theme.

⚠️ Important Premise

This LoRA is designed to be used natively on MOSS-Image (both MOSS-Image-SFT and MOSS-Image-Base). It is an extension, not a replacement. It works best when you leverage MOSS-Image's native strengths: JSON layout prompts, long prompt windows, and its strong text renderer.

Recommended base model: MOSS-Image-SFT for inference. Community evaluations consistently favour the SFT checkpoint over Base for LoRA inference, even when the adapter was trained against Base.

🔑 Trigger Words (Place at Beginning of Prompt)

Official V67 Moss Image Aesthetic trigger string:

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio

How to use it: place the full string at the very beginning of the prompt, before the scene description. It bundles three things — the model identity (Moss-Image), the text-rendering benchmark cue (LongLongTextBench (LLTB)), the commercial use-cases this component was tuned for (posters, menus, images for e-commerce and advertising materials), and the Moss ... aesthetic keyword family.

⚙️ Recommended Settings for MOSS-Image

ParameterRecommended ValueLoRA Weight0.78 – 0.90Base ModelMOSS-Image-SFT (photorealism & text) / MOSS-Image-Base (fidelity experiments)Steps50+ (≥ 50 recommended; text and fine texture degrade first if you cut steps)CFG / Guidance3.0 (useful range 3.0–6.0)SamplerEuler (default recommended)Shift5.0SLG (Skip Layer Guidance)Layer 22 uncond-skip (standard pipeline default)Resolution1M–2M total pixels (e.g. 1024×1024; 512–2048 px per edge, multiples of 16)Negative PromptEmpty or the light default set (see below)

These values come directly from the official MOSS-Image User Guide and community LoRA cards.

🎯 Starting Point Recommendations

GoalLoRA WeightCFGBase ModelSubtle aesthetic touch0.783.0SFTBalanced — Sweet Spot0.823.0SFTStrong style dominance0.883.0SFTPhotorealistic portrait focus0.853.5SFTArtistic / illustrative0.783.0BasePoster / menu / e-commerce layout0.78 – 0.823.0SFT

✨ What Makes This LoRA Special?

Enhanced Human Realism — Natural skin texture, individual hair strands, authentic expressions, no plastic look.
Natural Texture Fidelity — Sharp landscapes, detailed fur, flowing water, dense foliage.
Text-Preserving Aesthetic — MOSS-Image's dense-text rendering is its signature strength. This LoRA is tuned to keep that advantage, not blur it away.
Commercial Layout Ready — Explicitly tuned for posters, menus, images for e-commerce and advertising materials, where typography and composition must survive the aesthetic pass.
Layout-Compatible — Works with MOSS-Image's native JSON bounding-box prompts; the aesthetic enhancement does not interfere with spatial control.
Versatility — Generic subjects, no single-theme overfitting.

📊 Training Dataset Overview

Trained on 20 curated generic images across 5 categories (4 variations each), building a solid aesthetic foundation without overfitting:

CategoryDescriptionPortraitNatural skin texture, soft lighting, genuine expressionsLandscape & NatureForests, wildlife, detailed fur, morning lightArchitectureModern buildings, natural light, clean linesProductLuxury items, metallic reflections, macro detailsText-IntegratedPosters with precise typography and clean layouts

📸 Example Prompts (with MOSS-Image)

1. Architecture — Modern Building

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — a modern glass building in Shanghai, reflecting the sunset, clean architecture, minimalist, photorealistic, 8k, sharp details, natural textures.

Settings: MOSS-Image-SFT · LoRA 0.82 · CFG 3.0 · 50 steps · 1024×1024

2. Portrait — Natural Light

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — a young woman with natural skin texture, soft window light, individual hair strands, genuine smile, photorealistic portrait, 8k, sharp focus.

Settings: MOSS-Image-SFT · LoRA 0.85 · CFG 3.5 · 50 steps · 1024×1024

3. Wildlife — Forest Scene

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — a deer in a misty forest, detailed fur texture, morning light, natural foliage, cinematic landscape, photorealistic, national geographic style.

Settings: MOSS-Image-SFT · LoRA 0.82 · CFG 3.0 · 50 steps · 1024×1024

4. Product — Luxury Watch

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — luxury mechanical watch on leather, intricate dial, engravings, studio lighting, metallic reflections, product photography, macro, 8k.

Settings: MOSS-Image-SFT · LoRA 0.80 · CFG 3.0 · 50 steps · 1024×1024

5. Text-Integrated — Event Poster (with JSON Layout)

MOSS-Image's native JSON layout control combined with this LoRA is the headline use case: precise typography placement plus aesthetic quality. Keep the trigger inside the summary field so the aesthetic bias stays active alongside the layout instructions.

json

{
  "summary": "(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — minimalist event poster, geometric background, professional graphic design, high contrast, editorial quality",
  "items": [
    {
      "type": "text",
      "text": "MOSS-IMAGE CREATIVE ENGINE",
      "bbox": [80, 120, 920, 300]
    },
    {
      "type": "text",
      "text": "V67 AESTHETIC",
      "bbox": [80, 340, 600, 420]
    },
    {
      "type": "object",
      "description": "geometric abstract shapes, bold typography layout",
      "bbox": [80, 450, 920, 900]
    }
  ]
}

(Bounding boxes are illustrative — adjust to your canvas.)

Settings: MOSS-Image-SFT · LoRA 0.78 · CFG 3.0 · 50 steps · 1024×1024

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — elegant restaurant menu card, warm paper texture, fine serif typography, clean grid layout, soft natural light, editorial food photography inset, premium print finish, 8k.

Settings: MOSS-Image-SFT · LoRA 0.80 · CFG 3.0 · 50 steps · 1024×1024

7. E-commerce — Product on Clean Background

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio — premium skincare bottle on a soft neutral background, subtle shadow, label typography perfectly legible, commercial product photography, studio lighting, clean and minimal, 8k.

Settings: MOSS-Image-SFT · LoRA 0.80 · CFG 3.0 · 50 steps · 1024×1024

🚫 Negative Prompt

MOSS-Image's official guidance recommends a light, targeted negative set rather than a long laundry list:

text

nsfw, worst quality, bad anatomy, watermark, jpeg artifact

These are the tags the OpenMOSS team identifies as most effective. For aesthetic LoRAs specifically, community cards often recommend leaving the negative prompt empty — the model is not trained with a heavy negative branch, and a blank negative preserves prompt fidelity.

🧠 Advanced Tips for MOSS-Image

1. Use the JSON Layout Prompt

This is MOSS-Image's native superpower. Instead of hoping the model places text correctly, you tell it where text goes with bounding boxes. Layouts support controlled editing — change text, element descriptions, or box positions to adjust composition in a structured way. This is why the trigger string explicitly targets posters, menus, images for e-commerce and advertising materials.

2. Keep the Trigger Intact

Do not trim the trigger string. The first token block (Moss-Image, Caveat:1.1) anchors the LoRA, LongLongTextBench (LLTB) reinforces the text-rendering bias, the commercial use-case keywords steer composition toward layout work, and the Moss ... family carries the aesthetic. Removing parts of it measurably weakens the effect at lower LoRA weights.

3. Keep Steps at 50+

MOSS-Image degrades gracefully at low step counts on broad composition, but small text and fine textures are the first things to suffer. If your poster or menu has dense lettering, do not cut steps to save time.

4. CFG: 3.0 Is Usually Enough

The recommended range is 3.0–6.0. At 3.0, long prompts are handled well; short prompts may benefit from a small bump. Higher CFG can add contrast but also stiffness — the opposite of what an aesthetic LoRA is trying to achieve.

5. Long Prompts Are an Advantage

MOSS-Image accepts up to 4,096 tokens. Unlike many models where long prompts dilute attention, this architecture is built for it — use the room to describe texture, lighting, and layout intent in detail. The trigger string alone already consumes part of that budget, so write descriptions with intent.

6. Base vs. SFT

Use SFT for final renders — it produces clearer text grouping and more organised layouts. Use Base if you want rawer, more variable outputs or if you are training your own adapter.

7. ComfyUI Compatibility

If you run MOSS-Image in ComfyUI, use the native ComfyUI LoRA variant (key prefix diffusion_model.layers...), not the Diffusers/PEFT root file. Loading the wrong key namespace produces lora key not loaded warnings. Use the standard LoRA loader at model strength 1.0 and CLIP strength 0.0.

🔧 Troubleshooting

IssueSolutionEffect too weakIncrease LoRA weight to 0.88; make sure the full trigger string is present verbatim at the startPlastic / artificial lookAdd natural skin texture, photorealistic, natural detail; keep CFG at 3.0Text not rendering clearlyUse a JSON layout with explicit bounding boxes; keep steps at 50+; keep text strings simplePoster / menu layout collapsingLower LoRA weight to 0.78–0.80; move layout details into the JSON items array instead of the prose promptNature scenes look flatAdd Moss texture, Moss detail, fine detailArtifacts or distortionReduce CFG to 3.0; lower LoRA weight to 0.78Subject overfittingKeep the full trigger + keep descriptions genericLoRA not loading in ComfyUIUse the native ComfyUI key-format file, not the Diffusers root file

💪 Powerful Keywords

Quality: photorealistic, 8k, ultra-detailed, sharp focus, high detail, natural texture, fine detail, production-ready

Human: natural skin texture, individual hair strands, age-appropriate details, genuine expression, realistic portraits

Nature: detailed fur, foliage density, water flow, mist, reflections, landscapes, wildlife

Text: precise typography, accurate text rendering, clean layout, structured composition, JSON layout, bounding box

Commercial: posters, menus, images for e-commerce and advertising materials, catalogues, packaging mockups, label typography, print finish

Style: cinematic, editorial, architectural photography, commercial, minimalist, clean, balanced lighting

Brand: MOSS-Image, MOSS-Image-SFT, Moss image quality, Moss AI image, Moss diffusion, Moss text rendering

🙏 Philosophy & Credits

"MOSS-Image represents something rare in open-weight image generation: a model that is genuinely excellent at a specific, difficult task — dense, accurate text and spatial layout — rather than merely adequate at everything. This LoRA is designed to enhance and extend those capabilities across portraits, landscapes, products, posters, menus and e-commerce imagery, without erasing the model's native precision."
V67 Moss Image Aesthetic is built on a dataset of 20 curated images, offering a versatile foundation for professional, production-ready AI art, fully optimised for the MOSS-Image ecosystem.

🏢 About MOSS-Image / OpenMOSS

Released: 16 October 2026
Organisations: OpenMOSS + MOSI Intelligence, with the Shanghai Innovation Institute (SII), Fudan University, and Shanghai Jiao Tong University
Architecture: 5B single-stream DiT · frozen Qwen3.5-9B text encoder · FLUX.2 VAE
License: Apache 2.0 — weights and standalone pipeline code, no revenue threshold, no acceptable-use appendix, commercial use permitted
Variants: MOSS-Image-Base and MOSS-Image-SFT; community 8-step and 4-step distillations, GGUF and INT8 quantisations
Prompt window: up to 4,096 tokens
Layout control: native JSON bounding-box prompts for text and objects
Text rendering: 96.3% Mean-6 coverage on LongLongTextBench (self-reported)
Recognition: topped Hugging Face text-to-image trending shortly after launch; ~1.4k GitHub stars

🔮 Future & Updates

Base Model: MOSS-Image (Base / SFT)
Open Source: Yes — Apache 2.0, commercial use permitted
Next Steps: We are working on a v2.0 with a larger and more diverse dataset for even greater fidelity across portrait, product, and layout-heavy commercial domains.

🎨📷✨ V67 Moss Image Aesthetic — Native Precision for the MOSS-Image Engine

Download, test, and transform your ideas into reality. Don't forget to leave a Like — it is worth more than Buzz.

Trigger (copy-paste ready):

text

(Moss-Image, Caveat:1.1), LongLongTextBench (LLTB), posters, menus, images for e-commerce and advertising materials, Moss image quality, Moss AI image, Moss photorealism, Moss texture, Moss detail, Moss text rendering, Moss production, Moss diffusion, Moss vision, Moss cinematic, Moss studio

Files with SHA256: adb837fe6237cc2d2b73f3b371f8fdf2923cfff34ed14180585f11e824260ab6