One sheet per model. Each gives the formula that vendor publishes, whether the model takes a negative prompt at all, what the vendor explicitly tells you not to do, and a checklist to run before you send.
- Twelve models across image, video, text and audio.
- Four of them take no negative prompt at all — and on Runway the vendor documents that using one can produce the opposite.
- No two vendors publish the same formula. Google opens with the camera; ByteDance opens with format; everyone else opens with the subject.
- Each sheet is labelled by sourcing strength — vendor-documented, partly documented, or secondary. Three of the twelve are secondary and say so.
Which models take a negative prompt
The single most useful thing across all twelve sheets. Check yours before writing one.
On the models marked NO FIELD, every exclusion has to be rewritten as a description of what you want present. That is the vendors' own advice, not a workaround.
Image
FLUX.2
Black Forest Labs' own formula, the 30–80 word rule, and why a negative prompt does nothing here.
Black Forest Labs · Text-to-Image · VENDOR-DOCUMENTED · no negative prompt field
Midjourney v8.2
How --no actually behaves, the comma trap that splits your exclusions, and where the sourcing is weaker.
Midjourney · Text-to-Image · SECONDARY · negative prompt supported
ChatGPT Image
OpenAI publishes no prompt formula — but they do rewrite your prompt before generating it, and that changes how you iterate.
OpenAI · Text-to-Image · VENDOR-DOCUMENTED · no negative prompt field
Seedream 5.0 Pro
Write the exact words you want rendered, label all six references, and change one thing at a time.
ByteDance · Text-to-Image · VENDOR-DOCUMENTED · no negative prompt field
Video & motion
Seedance 2.5
ByteDance's eight-part template — format first, action third, and exclusions last in one Avoid clause.
ByteDance · Text-to-Video · VENDOR-DOCUMENTED · no negative prompt field
Veo 3.1
The one formula that opens with the camera — plus the audio syntax most guides never mention.
Google · Text-to-Video · VENDOR-DOCUMENTED · negative prompt supported
Runway Gen-4.5
Full sentences, not keywords — and the one model whose vendor documents that negatives can produce the opposite.
Runway · Text-to-Video · VENDOR-DOCUMENTED · no negative prompt field
Kling 3.0
The one video model with a real negative field — around 2,500 characters, and capacity is not permission.
Kuaishou · Text-to-Video · PARTLY DOCUMENTED · negative prompt supported
Higgsfield
Pick the preset and name the move in the prompt. Doing only one gets you approximately what you asked for.
Higgsfield · Video · Motion · SECONDARY · no negative prompt field
Text
ChatGPT · Claude · Gemini
What the three vendors actually publish — including the long-context rule worth up to 30 percent and why step-by-step is now conditional.
OpenAI · Anthropic · Google · Large Language Models · VENDOR-DOCUMENTED · negative prompt supported
Audio
ElevenLabs
There is no prompt box. The script is the prompt — and v3 and v2 take incompatible syntax.
ElevenLabs · Text-to-Speech · Voice · VENDOR-DOCUMENTED · no negative prompt field
Suno
Two fields, not one — and putting the wrong thing in the wrong box is why half your prompt gets ignored.
Suno · Music Generation · SECONDARY · no negative prompt field
Each sheet was written from the vendor's own published documentation where one exists. Where it does not — Midjourney, Suno, Higgsfield — the sheet says so at the top and the sourcing is labelled secondary rather than presented as vendor guidance. Kling sits between the two: its fields and limits are documented, the tuning advice is practice.
Everything here was checked on 25 August 2026. Model versions and vendor documentation both move; the negative-prompt table is the part most likely to change first. Corrections are logged at corrections.
For the same material as one continuous comparison rather than twelve sheets, see what the vendors actually say.