MOSS-Image-SFT_clear Released! An Open Model Trained on 3 Billion Images with Strong Potential for Anime Art

Introduction

Hello, this is Easygoing.

Today I’m introducing MOSS-Image-SFT_clear, a fine-tuned version of the image generation AI MOSS-Image (trained on 3 billion images) that shows strong potential for anime-style illustrations.

OpenMOSS Is an Industry-Academia Collaboration Project

MOSS-Image is an image generation AI model released on October 16, 2026 by OpenMOSS, an organization based in Shanghai, China.

OpenMOSS is a research project centered on Fudan University in Shanghai, with participation from Shanghai Innovation Institute and the startup MOSI Intelligence. Its goal is to accelerate open-source AI research through industry-academia collaboration.

MOSS-Image Uses a Modern Architecture

MOSS-Image consists of the following three components:

MOSS-Image uses Alibaba’s Qwen-3.5 (9B) as the text encoder, MOSS-Image-Transformer as the transformer, and Black Forest Labs’ Flux.2-vae-small as the VAE.

All components are licensed under the open Apache-2.0 license, so both the model and its outputs can be used commercially with confidence.

MOSS-Image Excels at Prompt Understanding

Among recently released image generation models, MOSS-Image is relatively lightweight, yet about two-thirds of its size is taken up by the text encoder that analyzes prompts.

The text encoder is the large 9B-parameter Qwen-3.5 model released in March 2026. It has excellent prompt understanding and is also used as the prompt enhancer, allowing efficient execution within the same workflow.

On the other hand, the transformer is only about 10 GB—relatively light for recent models—so outputs from MOSS-Image tend to show more detail breakdowns and weaker three-dimensional rendering compared with high-quality models such as Krea 2.

MOSS-Image tends to produce more artifacts in details such as fingers.

Still, by offloading the text encoder to RAM, the transformer can be fully loaded into 12 GB of VRAM, giving it a speed advantage especially on mid-range GPUs.

MOSS-Image-Base Was Trained on 3 Billion Images!

Two versions of MOSS-Image have been released:

• MOSS-Image-Base: The base model used for training
• MOSS-Image-SFT: A further fine-tuned model optimized for generation

MOSS-Image-Base is a base model trained on approximately 3 billion images, primarily large open-source datasets such as Re-LAION-5B, OmniCorpus, and DataComp. It can generate a rich variety of illustrations ranging from anime to photorealistic styles.

MOSS-Image-SFT is a re-tuned version of the base model designed for more stable outputs. While the variety of styles is slightly reduced, overall quality is higher, so it is recommended for regular generation.

Introducing MOSS-Image-SFT_clear!

I fine-tuned the MOSS-Image-SFT model toward clearer, more anime-oriented outputs and created the MOSS-Image-SFT_clear model.

Below is a comparison between MOSS-Image-SFT_clear and the original MOSS-Image-SFT.

Captain

Agent

Compared with the original MOSS-Image-SFT on the right, MOSS-Image-SFT_clear produces cleaner, less noisy results with improved detail expression.

Saturation and contrast are also higher, resulting in clearer, more convincing illustrations.

More Examples!

Here are additional examples generated with MOSS-Image-SFT_clear.

All prompts were created by extending the system prompt of MOSS-Image’s official prompt enhancer and captioning the images.

Terminal Station

Anime-style digital illustration of a young woman standing in a futuristic, neon-lit transit station. She has short, dark hair with a single curl at the crown and large, amber-colored eyes that gaze directly at the viewer. Her right hand is raised, with her index finger gently touching her lower lip in a contemplative gesture. She wears a dark grey, hooded utility jacket with a white collar and a purple scarf peeking out from underneath. On the left side of her jacket is a blue rectangular patch displaying abstract graphics, and on the right sleeve is a purple circular emblem. The background is a dense, multi-level urban environment filled with glowing yellow and orange signage, digital information boards, and other commuters walking along the platforms. The floor is highly reflective, mirroring the warm ambient light from the overhead signs and the cool blue tones of the station's architecture. The scene is illuminated by a mix of hard, directional light from the signs and soft, diffuse glow from the interior, creating a vibrant, high-contrast atmosphere typical of anime key visuals. The composition places the subject centrally, with a shallow depth of field blurring the distant crowd and signage to emphasize her presence in the bustling, cybernetic cityscape.

Sunset Cobblestone Street

Anime-style digital illustration of a young girl with short, tousled brown hair and large, expressive blue eyes, smiling softly as she glances back over her shoulder. She is dressed in a vibrant royal blue dress with white ruffled cuffs and a high collar, cinched at the waist with a brown leather belt. Her pose is casual and inviting, with her upper body turned toward the viewer while her gaze remains engaged and warm. The setting is a charming medieval European town bathed in the golden glow of a sunset. She stands on a cobblestone street that curves gently into the distance, flanked by timber-framed houses with glowing windows. In the background, a stone arch bridge spans a river or valley, silhouetted against a sky filled with soft clouds tinged in purple and orange. Small figures of people can be seen walking along the distant path, adding life to the scene. The composition uses a shallow depth of field, keeping the girl in sharp focus while the town and sky fade softly into the background. Warm ambient light from the setting sun illuminates the edges of her hair and the cobblestones, creating a glowing, dreamlike atmosphere. The overall palette is rich with warm yellows and oranges contrasting against the cool blue of her dress, evoking a sense of nostalgia and tranquility. The art style features clean lines, stylized facial features, and vibrant, saturated colors typical of modern anime key visuals.

Gaze

Japanese anime key visual of a stern geisha in a traditional kimono, rendered with high-contrast cel-shaded lines and flat, vibrant color fields. The subject is a young woman with pale, porcelain skin and a severe, unsmiling expression, her gaze locked directly on the viewer. Her hair is styled in a formal, high bun secured with two golden hairpins, with a few fine, dark strands framing her sharp jawline. She wears a heavy, deep indigo kimono with broad, sweeping sleeves that dominate the foreground, accented by a wide, ornate purple obi tied at the waist with intricate floral embroidery in red, green, and gold. A small, golden crest is visible on the sleeve of her garment. Her makeup is precise and dramatic, featuring heavy, winged eyeliner, soft pink blush on her cheeks, and a muted red mouth. The lighting is sharp and directional, casting deep, angular shadows that emphasize the folds of the fabric and the structure of her face. In the background, a blurred, impressionistic landscape features a pale sky with soft clouds and the dark silhouette of a pine tree, suggesting a traditional Japanese garden or veranda. The composition is a tight close-up, focusing entirely on her upper body and face, creating an intimate yet commanding presence. The art style is distinctly modern anime, characterized by bold outlines, saturated colors, and a graphic quality reminiscent of a printed poster or manga cover.

Fortune-Telling Room

A cinematic portrait of a woman in a shimmering silver saree, captured with a photorealistic style and a dynamic Dutch angle. She is framed in a tight close-up, focusing heavily on her face, which bears a subtle, confident smile. Her dark, wavy hair cascades over her right shoulder, contrasting with her fair skin and heavy, smoky eye makeup. She wears elegant dangling earrings that catch the light. Her attire consists of a floor-length saree woven entirely from silver sequins, creating a textured, sparkling surface that reflects the ambient light. The drape is sophisticated, with the fabric falling smoothly across her torso and shoulder. She is seated, her posture relaxed yet poised, with a glimpse of a silver candlestick holder featuring a thick, textured orange candle visible to her left in the midground. The background features a warm, wood-paneled wall adorned with framed black-and-white photography, adding depth and a gallery-like atmosphere to the scene. The lighting is dramatic and moody, with a strong key light illuminating her face and the silver fabric, creating sharp highlights and deep shadows that emphasize the texture of the sequins and the contours of her features. The perspective is tilted, lending a sense of movement and modern energy to the composition. The overall tone is sophisticated, glamorous, and intimate, capturing a moment of quiet elegance.

ComfyUI Workflows

Here are ComfyUI workflows for using the MOSS-Image-SFT_clear model.

1. Simple Illustration Generation

• MOSS-Image-SFT_clear_simple_20261031.json

Models (placement folders)

• diffusion_models (ComfyUI/models/diffusion_models/)
  • MOSS-Image-SFT_clear_BF16
  • MOSS-Image-SFT_clear_Int8_ConvRot_HQ (lightweight version)
• text_encoders (ComfyUI/models/text_encoders/)
  • qwen3.5_9b_bf16.safetensors
  • qwen3.5_9b_int8_convrot.safetensors (lightweight version)
• vae (ComfyUI/models/vae/)
  • flux2-vae_small.safetensors

This is a simple workflow for generating illustrations with the MOSS-Image-SFT_clear model.

Compared with the official ComfyUI workflow, the Negative Prompt is left empty and the SkipLayerGuidanceDiTSimple node is removed to suppress model-specific noise and achieve cleaner outputs.

About empty negative prompts

Note that the SkipLayerGuidanceDiTSimple node skips specified blocks only during unconditioning calculation. In the original ComfyUI workflow it functions similarly to a text-free negative prompt.

2. Image-to-Prompt Generation + AI Upscaling

• MOSS-Image-SFT_clear_image_prompt_upscale_20261031.json

Models (placement folders)

• diffusion_models (ComfyUI/models/diffusion_models/)
  • MOSS-Image-SFT_clear_BF16
  • MOSS-Image-SFT_clear_Int8_ConvRot_HQ (lightweight version)
• text_encoders (ComfyUI/models/text_encoders/)
  • qwen3.5_9b_bf16.safetensors
  • qwen3.5_9b_int8_convrot.safetensors (lightweight version)
• vae (ComfyUI/models/vae/)
  • flux2-vae_small.safetensors
• upscale_models (ComfyUI/models/upscale_models/)
  • 4x_IllustrationJaNai_V3detail_FDAT_M_40k_fp16.safetensors (Non-commercial model)

This workflow uses the Qwen-3.5 (9B) text encoder from MOSS-Image to generate prompts via image captioning, then further refines the result with AI upscaling.

With this workflow you can simply feed any illustration to generate a prompt optimized for MOSS-Image.

Summary: Try MOSS-Image-SFT_clear!

• MOSS-Image was trained on 3 billion images
• It has some detail artifacts but offers rich variety
• Released under the open Apache-2.0 license

In this post I introduced the MOSS-Image-SFT_clear model.

Although MOSS-Image produces more detail breakdowns than some models, it offers rich output variety and excellent characteristics as a base model. It also has a significant advantage over existing models in terms of licensing.

One of the core developers of MOSS-Image has stated on a Reddit thread that releasing an anime-oriented model is a priority, so we can look forward to derivative models with additional training or fine-tuning specifically for anime illustrations.

With MOSS-Image-SFT_clear you can generate quite clear and stable illustrations, so why not give it a try?

Thank you for reading to the end!