mitsua-diffusion-one-GGUF-Q8_0

Mitsua Diffusion One — GGUF Q8_0

This repository provides a GGUF Q8_0 conversion of Mitsua Diffusion One for use with stable-diffusion.cpp and compatible local applications.

Modified distribution notice: The original Mitsua model weights were converted and quantized to GGUF Q8_0. This repository does not claim new training or authorship of the underlying model. Please use the original Mitsua repository for the authoritative model card, training information, and creator examples.

Mitsua Diffusion One was developed by Abstract Engine. Its original model card states that it was trained from scratch using public-domain/CC0 images or copyrighted images used with permission, with a fixed OpenCLIP text encoder.

License and required notice

Copyright (c) 2023 Abstract Engine Co., Ltd. and contributors.

The model and this quantized derivative are distributed under the Mitsua Open RAIL-M License. Read the complete local license copy and the license in the original repository before downloading or using the model.

By downloading, accessing, using, reproducing, or distributing this model or a derivative, you agree to comply with the complete Mitsua Open RAIL-M License, including all use-based restrictions in paragraph 5 and Attachment A.

The license contains enforceable use-based restrictions. Among other requirements, users must not:

• use the model for unlawful or harmful purposes listed in the license;
• infringe the rights of others by supplying image sources or model weights without the necessary rights or permission;
• misrepresent generated images as not AI-generated.

Users are responsible for their inputs, outputs, and compliance with the complete license. This summary does not replace the license and is not legal advice.

Files

Verify the download before use:

0863ed89a5aa2881d8039cfeb6b29d55d4e2ca67819ffbd74558a6b639f5fd5d  mitsua-diffusion-one-q8_0.gguf

Compatibility

This file is intended for stable-diffusion.cpp, not direct loading through Diffusers or Transformers.

Verified configuration:

• Runtime: stable-diffusion.cpp master-820-de298c2 (de298c2)
• Generation modes validated in this model card: text-to-image
• Sampler used for the examples: ddim_trailing (creator setting: DDIM)
• Scheduler: simple
• Prediction mode detected by the runtime: SD 2.x, epsilon prediction
• Backends tested: Android OpenCL on Qualcomm Adreno 830

Compatibility with other stable-diffusion.cpp revisions, devices, backends, or applications may vary. Image-to-image is supported by stable-diffusion.cpp, but the examples below validate text-to-image only.

stable-diffusion.cpp example

The creator-style DDIM settings map to ddim_trailing with the simple scheduler in stable-diffusion.cpp:

sd-cli \
  --model mitsua-diffusion-one-q8_0.gguf \
  --prompt "beautiful huge waterfalls, surrounded by msssive cherryblossom, a post impressionism painting" \
  --negative-prompt "photo," \
  --width 768 \
  --height 512 \
  --steps 40 \
  --seed 3261478676 \
  --cfg-scale 7.5 \
  --sampling-method ddim_trailing \
  --scheduler simple \
  --output waterfall.png

Backend selection and memory options depend on how stable-diffusion.cpp was built and on the target device.

Mobile GGUF showcase

The following images were generated locally on a Samsung Galaxy S25 using the GGUF file in this repository, stable-diffusion.cpp, and the Qualcomm Adreno 830 OpenCL backend. They are AI-generated images.

The prompts, seeds, dimensions, DDIM sampler, CFG scale, step count, and negative prompts reproduce selected entries from the original creator's prompt and settings file. Compare them with the original Mitsua model card and showcase.

<Gallery />

Reproduction settings

Waterfall and cherry blossoms

beautiful huge waterfalls, surrounded by msssive cherryblossom, a post impressionism painting

Old man portrait

old man portrait, an impressionism painting, fine art,  white hair, glasses

Green train

green color train at railway station, neon, illumination, lighitng, night, vignette in film, black backgroud, futuristic, cinematic lighting

Android test observations

• VAE spatial tiling was enabled for the 512×768 and 768×512 generations.
• The model used approximately 1.98 GB of GPU-backed parameter memory in the tested runtime.
• All three 40-step generations completed without reaching Android's SEVERE thermal status.
• External fan cooling was used. Long generations should begin only after the device has cooled, and performance varies significantly with starting temperature and device thermal policy.
• The timings above are observations from one device and are not general benchmarks.

Limitations

• Quantization can cause numerical and visual differences from the original weights.
• Results may differ between CPU, OpenCL, Vulkan, CUDA, Metal, and other backends even with identical prompts and seeds.
• Prompt interpretation, anatomy, composition, and fine details may be inconsistent.
• This repository does not provide hosted inference and does not guarantee compatibility with Hugging Face Inference Providers.
• Users must evaluate whether their intended inputs and uses comply with the Mitsua Open RAIL-M License.

Original project and credits

• Original model: Mitsua/mitsua-diffusion-one
• Original prompts and settings: mitsua-diffusion-one-prompts.csv
• Original license: Mitsua Open RAIL-M License
• Original developer: Abstract Engine / Mitsua
• Local runtime: stable-diffusion.cpp

Thank you to Abstract Engine, Mitsua contributors, and the upstream open-source projects that made this local conversion and mobile testing possible.