| 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 Prompting Guide
What SD-Turbo is
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.
Key differences from standard SD
| 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 |
Prompting rules
1. Write dense, front-loaded prompts
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)
2. Skip negative prompts entirely
They have no effect at one step. Don't waste tokens on them.
3. Use concrete visual descriptors
- Subject first, then style/mood
close take, top view, wide shot for framingcinematic, photo, detailed for style- Color/lighting descriptors work well:
warm lighting, golden hour
4. Resolution
512×512 is the preferred resolution. Higher works but quality degrades.
Best use cases
- Live preview loops — iterating on a prompt with instant visual feedback
- Rapid prototyping — quick concept exploration before a slower high-quality pass
- Research — real-time generative model experiments
What SD-Turbo is NOT good at
- Fine detail (hair, text, texture) — softer than full SD
- Complex prompt adherence — distilled model loses some nuance
- Production-quality final outputs — use SDXL for that
Future direction
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.
Model
Generating an image
python3 scripts/tokenize.py models/sd-turbo/tokenizer/vocab.json models/sd-turbo/tokenizer/merges.txt "your prompt here" /tmp/tokens.txt
rcc ordi-jla /tmp/tokens.txt M:/machin-diffusion/tokens.txt 30
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
