Which controls move output away from the default?
What each control does, in the words of the paper or documentation that defines it.
| Control | Where it is defined | What the source says |
|---|---|---|
| Guidance scale | Hugging Face diffusers, Stable Diffusion text-to-image pipeline | Defaults to 7.5. A higher value ties the image more closely to the prompt "at the expense of lower image quality"; guidance is on when the value is above 1 |
| Classifier-free guidance | Ho and Salimans (2022) | Combines conditional and unconditional estimates "to attain a trade-off between sample quality and diversity" |
| Spatial conditioning (ControlNet) | Zhang and others (2023) | Adds conditions such as edges, depth, segmentation and human pose to a large pretrained diffusion model; training was robust with small (under 50,000) and large (over 1 million) datasets |
| Negative prompt | Hugging Face diffusers | The prompt that guides "what to not include in image generation"; ignored when guidance is off |
Hosted generators expose fewer of these. Where a tool lets you set a reference image, a pose, a depth map or an edge sketch, that is spatial conditioning by another name.
What should you add back?
These are working methods. None of the sources above tests them:
- Grain and flare. Optical noise breaks up the even smoothness the origins page traces to filtering for appeal.
- Imperfect subjects. Real faces, lived-in rooms and bodies mid-motion, instead of the idealized version a prompt returns by default.
- Constraint pipelines. Depth, pose and edge conditioning, or a locked style reference, so the sampler cannot drift to its average.
- Mixed media. A generated pass under photography, film grain or compression, so no single source sets the look.
Sources
Checked 13 November 2026:
- Hugging Face, Diffusers: Text-to-image pipeline (guidance_scale and negative_prompt)
- Ho, J. and Salimans, T. (2022), Classifier-Free Diffusion Guidance, arXiv:2207.12598
- Zhang, L. and others (2023), Adding Conditional Control to Text-to-Image Diffusion Models, arXiv:2302.05543
Next step
Check the result against the tells on the signature page, and plan the final size with the print size planner.