All articles

Guides

Why Negative Prompts Do Nothing in Turbo Diffusion Models

Negative prompts depend on classifier-free guidance, and few-step turbo models run without it. Why that field is ignored, and what to do instead.

By · · Updated · 6 min read

Negative prompts do nothing in turbo diffusion models because they depend on classifier-free guidance, and few-step turbo models run without it. On classic diffusion models they work. Ask an experienced Stable Diffusion user for advice and you will probably get one: a comma-separated list like “blurry, deformed, extra fingers, watermark, low quality” meant to ward off every known way an image can go wrong. On the turbo models that make on-device generation possible, that list is ignored.

That follows from how few-step models buy their speed, and the same idea shows what to do instead.

What classifier-free guidance is

A diffusion model builds a picture by guessing the noise in a half-formed image and subtracting it. Then it does the same thing again, and again. Each round is called a denoising step. Left alone, that process follows your prompt only loosely. One trick makes your prompt bite. (The technical name for it is classifier-free guidance, or CFG.) At each step, the model guesses twice instead of once:

  1. a guess with your words in mind (the conditional prediction): what the noise looks like given your prompt, and
  2. a second guess made without your words (the unconditional prediction): what the noise looks like given no prompt at all, or given the negative prompt instead.

The sampler then measures the gap between those two guesses and pushes the image along it. How hard it pushes depends on the guidance value, the “cfg scale” you may have seen set to 7 or so in desktop tools. The image gets steered toward what the positive prompt describes and away from that second guess. A negative prompt works only because it takes over that second slot. Write “blurry” there and every step shoves the image away from the model’s idea of blurriness.

A negative prompt is one input to a steering system that needs two guesses. Take away the system and there is nothing to steer.

How hard you steer is itself a trade-off, which is why desktop users tune it. Set it too low and the prompt barely registers. Set it too high and images turn oversaturated and burned, because the steering overpowers the cleanup work. The 6–8 sweet spot everyone converged on is a compromise. Turbo models set out to make that compromise unnecessary.

Why turbo models turn it off

Guidance has a price. Two guesses per step means running the diffusion backbone twice per step. That is double the compute for every image. On a server farm the cost is accepted. On a phone it would roughly double generation time and battery drain.

Turbo models (Stable Diffusion Turbo, SDXL Turbo, and their relatives) are trained as fast copies of a slower model. A student network learns to do in one to four steps what the original needed dozens of steps to do. (The name for this training is distillation, and it uses techniques like adversarial diffusion distillation.) Part of that training bakes the prompt-following behavior straight into the single conditional guess. So the finished model is built to run at a guidance scale of zero, or one, depending on the convention, with one guess per step and no second pass.

That design is why a 1–4 step model is fast enough for a phone. But once the second guess is gone, so is the slot the negative prompt used to fill. There is nowhere in the computation for those words to go.

The placebo text box

On a turbo model, a negative prompt field is a placebo. Type “ugly, deformed, watermark” or type a grocery list. Given the same seed, the output bits are identical, pixel for pixel. Some interfaces still show the field so they stay compatible with non-turbo models. That is how the folklore survives. People fill it in, get a decent image, and credit the incantation.

This applies to models distilled to run without guidance. Classic Stable Diffusion at 20–50 steps with cfg around 7 honors negative prompts fully. Some few-step methods keep partial guidance at low scales, and negatives have a weak effect there. For SD Turbo and SDXL Turbo in their documented guidance-free configuration, the negative prompt has no effect.

Read the model documentation and the app’s settings. A missing guidance slider does not prove that guidance is disabled: it may be fixed internally. In a tool with controlled seeds, comparing negative prompts can help, but keep every other setting identical.

What to do instead

Everything a negative prompt did can be done another way, usually better.

  • State the positive opposite. Instead of “blurry” in a negative box, put “sharp focus, fine detail” in the prompt. Instead of “cartoon,” write “photograph, 35mm, natural light.” Turbo models follow concrete positive styling well. Our prompting guide for on-device models collects the patterns that work.
  • Change the seed. Most of what people fight with negative prompts is seed luck. A weird pose or a muddled background usually comes down to the roll. Rerolling is cheap when an image takes seconds.
  • Use img2img for global steering. If an image is close but the style is off, feed it back through image-to-image with a revised prompt at moderate strength. You keep the composition and re-decide the style.
  • Use inpainting for local fixes. A negative prompt was always a crude tool for “fix this one region.” Masked inpainting is the precise one. You paint over the offending area and regenerate only that part, leaving the rest untouched.

How Onhand handles this

Onhand does not expose a negative-prompt box, a CFG slider, or manual seed locking. Turbo and Detailed Images use the guidance-free behavior described here. Other models in the library use different internal settings; some Core ML style models have guidance and a preset negative prompt.

The controls available to you are the positive prompt, style and quality choices, editing strength, and the supported editing tools. Use Make variations for new attempts, Refine to work from a result, or Turbo inpainting for a selected area. Generation runs on the device.

To change one area instead of steering the whole image away from something, see AI inpainting on iPhone. Onhand vs Draw Things covers the same Turbo limit next to an app with a much larger set of generation settings.

Sources

Primary sources for the claims above, so you can check them rather than take our word for it.