Unofficial community derivative. Not an OpenMOSS release. This page is a mirror/share only.

An unofficial, all-in-one ComfyUI checkpoint for anime-oriented image generation based on MOSS-Image. It combines an INT8 ConvRot, fused-QKV diffusion backbone with the native Qwen3.5-9B text encoder and FLUX.2 VAE into a single file. The diffusion backbone incorporates the community miaoka turbo1 derivative and an NSFW LoRA at 0.7 strength. It is biased toward anime illustration and adult-content capability compared with the official SFT model.

License: Apache-2.0
Base: OpenMOSS-Team/MOSS-Image-SFT
Author of this merge: Langzaigg (Hugging Face)
This upload: Mirror/share only. Not my checkpoint.

Architecture & Components

• Diffusion backbone: INT8 ConvRot with fused-QKV layout

• Text encoder: Native Qwen3.5-9B (included)

• VAE: FLUX.2 VAE (included)

• Merged components:

  • Community miaoka turbo1 derivative (in the diffusion backbone)

  • NSFW LoRA at 0.7 strength

• Bias: Anime illustration + adult content capability

No separate text encoder, VAE, LoRA, or QKV-remapping patch is required.

File

FileSizeFormatmoss_image_anime_turbo.safetensors14.44 GiB Safetensors checkpoint

Place the file in:

ComfyUI/models/checkpoints/

Load with CheckpointLoaderSimple or a similar node.

Recommended Settings

This is a turbo model. Do not treat it like a standard 20–50 step diffusion model.

Steps 4

Samplerres_multistep

Schedulersimple

CFG1.0

Negative prompt

Empty

Denoising1.0

Resolution~1 MP (width and height divisible by 16)

Showcase uses CFG 1.1 WITH negative prompts and 2 MP resolution, so you see this works as well.

Fallback: If res_multistep produces anomalies, use euler + simple at 4 steps.

Usage

1. Download moss_image_anime_turbo.safetensors.

2. Place it in ComfyUI/models/checkpoints/.

3. Load via CheckpointLoaderSimple.

4. Use the settings above.

5. Optional: Works with ComfyUI-AnimaFlow for structured tags + English descriptions and HTTP/MCP workflows.

Input Support

• Natural-language captions

• MOSS Layout JSON (positioned elements, panels, text)

• Reference images (semantic guidance only; does not preserve identity, clothing, or logos exactly)

Evaluation Notes (from original HF card)

• 95/95 successful generations in controlled tests (19 scenes, same prompts/seeds/resolutions).

• Tested with Euler (4/6 steps) and res_multistep + simple (4 steps).

• Updated checkpoint showed improved backgrounds, clothing, materials, and complex actions compared with an earlier version.

• Median generation time (updated checkpoint, 4 steps): ~27–28 seconds depending on sampler.

Limitations

May fail at:

• Complex anatomy

• Object interactions

• Exact layout adherence

• Negative-concept suppression

Small or multi-paragraph text may be malformed. Prefer short literal strings.
Reference images provide semantic guidance only.

Safety

This checkpoint contains an NSFW LoRA and can generate adult content.

• Do not generate sexual content involving minors.

• Avoid non-consensual intimate imagery or harmful content.

• Use only where lawful.

For safer outputs, explicitly specify clothing, pose, and scene in the prompt.

Credits

• Base model: OpenMOSS-Team/MOSS-Image-SFT (Apache-2.0)

• INT8 ConvRot source: toxicdog/MOSS-Image-SFT-Turbo-INT8-ConvRot

• Community components: miaoka turbo1 derivative and NSFW LoRA (merged)

• This merge: Langzaigg – Hugging Face

• This CivitAI page: Mirror/share only. All credit to the original authors and contributors listed above.

License

Apache-2.0 for this model and upstream components. Verify individual licenses of any further derivatives before redistribution or production use.

This is an unofficial merge. Verify provenance and comply with upstream terms.