メニュー
SD-Turbo prompting guide. How to write prompts that work with 1-step distilled generation (no CFG, no negative prompts). Read this before generating images.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/javimosch/machin-diffusion --skill machin-diffusion-promptingコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Optimization history (434s → 19.8s), patterns that worked, and the roadmap for further gains. Read this before planning optimizations.
Architecture and pipeline overview for machin-diffusion (SD-Turbo in pure MFL). Read this first to understand the codebase before making changes.
OpenCL GPU kernel design patterns, caveats, and how to add new builtins to the machin runtime. Read this before optimizing GPU kernels or adding new builtins.
| name | machin-diffusion-prompting |
| description | SD-Turbo prompting guide. How to write prompts that work with 1-step distilled generation (no CFG, no negative prompts). Read this before generating images. |
SD-Turbo is a distilled version of Stable Diffusion 2.1, built for single-step, real-time text-to-image generation instead of the usual 20-50 step diffusion process. Stability AI labels it a "research artifact" rather than a production model — SDXL-Turbo is recommended for better prompt understanding and image quality.
Think of it as a latency optimization: the tradeoff is speed for fidelity and prompt adherence.
| Aspect | SD-Turbo | Standard SD/SDXL |
|---|---|---|
| Steps needed | 1 (single network eval) | 20–50 |
| Guidance scale | Not used — guidance_scale=0.0 | 7–12 typical |
| Negative prompts | Not effective at 1 step | Fully supported |
| Resolution | 512×512 preferred | Up to 1024×1024 |
| Best for | Real-time previews, rapid prototyping, research | Final, detailed outputs |
Since there's no CFG steering, the model relies purely on how well the prompt tokens map to its distilled understanding. Put the most important descriptors first.
Good: a girl photo, close take
Good: planet earth seen from the moon, detailed, cinematic
Bad: a photo of a girl who is standing close to the camera and looking at the viewer (too verbose, key info buried)
They have no effect at one step. Don't waste tokens on them.
close take, top view, wide shot for framingcinematic, photo, detailed for stylewarm lighting, golden hour512×512 is the preferred resolution. Higher works but quality degrades.
Newer turbo variants (2025-2026) use Adversarial Diffusion Distillation (ADD) for 4-step generation with better quality retention. SD3.5 Large Turbo is an example. These may be better fits if you want speed without sacrificing as much prompt fidelity.
models/sd-turbo/# 1. Tokenize
python3 scripts/tokenize.py models/sd-turbo/tokenizer/vocab.json models/sd-turbo/tokenizer/merges.txt "your prompt here" /tmp/tokens.txt
# 2. Deploy tokens
rcc ordi-jla /tmp/tokens.txt M:/machin-diffusion/tokens.txt 30
# 3. Run (uses seed 12345 by default)
rcx ordi-jla "cmd /c M:\machin-diffusion\machin-diffusion.exe M:\machin-diffusion\models\sd-turbo M:\machin-diffusion\tokens.txt M:\machin-diffusion\output.ppm" 600