Negative Prompt for Realistic People: What to Exclude, Why

A good negative prompt for realistic people is short and specific: "plastic skin, airbrushed, doll eyes, extra fingers, deformed hands, cartoon, 3d render, watermark" covers the common failures in Stable Diffusion workflows. Each term should target a problem you actually see, and in Flux, which does not support negative prompts, you write the same intent as a positive description instead.

Checked October 2026 against Black Forest Labs' guide to working without negative prompts, ComfyUI's text-to-image tutorial and Midjourney's --no parameter. How strongly each term works varies by model, checkpoint and settings.

Most negative prompts online are copied lists of fifty terms with no explanation. This page does the opposite: it explains what each common term is meant to suppress, which model families use negative prompts at all, and how to build a short list that fits your persona. For the full image pipeline, start with our AI image generation for influencers guide.

Negative prompt for realistic people: which models use one

In ComfyUI, the negative prompt is a separate text encode node wired to the KSampler's negative input, alongside the positive one. Black Forest Labs is explicit that FLUX models do not support negative prompts, and adds that models generally struggle with negation: writing "a person without glasses" puts attention on glasses. That is the most important point on this page if you work in Flux.

Negative prompt for a realistic portrait: what each term does

Read the right-hand column as the more reliable fix in every model. Negative terms push away from a failure; positive descriptions pull toward what you want. In Stable Diffusion workflows, use both. In Flux, use only the right-hand column.

Stable Diffusion negative prompt for a person: the template

Positive:
photo of [persona], [age range], [hair], natural skin texture,
visible pores, soft window light, 85mm lens, f/1.8,
candid expression, realistic photograph

Negative (Stable Diffusion / SDXL):
plastic skin, airbrushed, doll eyes, extra fingers,
deformed hands, cartoon, 3d render, oversaturated,
blurry, watermark, text

Flux negative prompt: what to do instead

Because Flux has no negative prompt, the same goals have to be stated positively. Black Forest Labs' guidance is to describe what you want, not what you do not want. The template below turns the Stable Diffusion negative list into positive descriptions.

Flux (no negative prompt - say it positively):
photo of [persona], [age range], [hair], natural skin
with visible pores and slight imperfections, relaxed hands
resting on a cafe table, soft overcast daylight, muted
natural colours, shot on an 85mm lens, sharp focus on the eyes

Midjourney: the --no parameter

Midjourney's documentation describes --no as a way to tell it what you do not want, followed by the elements to leave out, for example --no text, watermark. Keep it short. For realistic people, most of the effect still comes from the main prompt: light, lens and skin texture. Our guide to camera settings for AI photos covers those terms in depth.

Building your own list, step by step

A negative prompt that works for your persona is something you build, not something you download. The process is the same whatever checkpoint you use, and it takes about an hour the first time.

1. Fix the seed and settings. Use one seed, one sampler and one CFG value for the whole test, so the only thing that changes is the negative prompt.
2. Generate a baseline with an empty negative. Save it. This is what the model does on its own with your positive prompt.
3. Name the failures. Look at the baseline at full size and write down what is wrong in plain words: skin too smooth, eyes too large, an extra finger, a painterly look.
4. Add one term per failure. Use the table above to pick the closest term. Regenerate with the same seed and compare.
5. Keep only terms that help. If a term changes nothing, or makes something else worse, remove it.
6. Test on new seeds. A list that fixes one image may not generalise. Run five or six new seeds before calling it done.

Keep the final list with a note of the model and settings it was built for. When you switch checkpoints or move to a new model family, repeat the test rather than assuming the list carries over.

When a negative prompt is the wrong tool

Some problems are better fixed elsewhere in the workflow. Hands and fingers often improve more from a simpler pose, a crop that keeps hands out of frame, or targeted inpainting than from any negative term. Faces that drift between images are a consistency problem, solved with a trained character LoRA or reference-image methods rather than prompt wording. And skin that looks smooth after upscaling may come from the upscaler or a face-restoration step, not from the prompt at all. If a negative term is not fixing a problem after a few tries, look at the rest of the pipeline before adding more words.

Mistakes that make negative prompts worse

• Pasting a hundred-term list. Each term gets less weight, and contradictory terms fight each other.
• Negating what you want to see. Terms like "beautiful" or "young" in a negative list can distort faces in unexpected ways.
• Fixing anatomy with words only. Hands improve more from simple poses and inpainting than from "extra fingers" in the negative.
• Reusing one list across models. A list tuned for one checkpoint can hurt another.
• Using negation in Flux. It can add the very thing you name.

Negative prompts help with individual images; consistency across a whole feed comes from a trained character and a fixed workflow. Our LoRA training guide covers the face, and our AI Influencers program covers the full persona system from prompts to publishing.

Negative prompts for realistic people: FAQ