# Negative prompts and CFG scale explained, with practical lists > What a negative prompt removes, what CFG scale controls, which image models and video engines honor each, and copy-paste lists for photoreal and anime work. Published 2026-09-08 ยท eroq.ai โ€” canonical: https://eroq.ai/blog/negative-prompts-explained Two controls sit under every image prompt in the Image studio, and most people never touch them. That is a shame, because between them they answer the two most common complaints about generated images: "it added something I did not ask for" and "it ignored the thing I did ask for". This is what a negative prompt and a CFG scale each do, where they apply on images and on video, and lists you can paste in and forget. ## What a negative prompt does A diffusion model builds an image by steering noise toward your prompt, step by step. The [negative prompt](/glossary/negative-prompt) is a second direction it steers *away* from. It is a push, not a filter: listing "hat" makes a hat less likely, not impossible, and listing "ugly" does nothing measurable because the model has no single idea of ugly to avoid. The negative prompt earns its keep on three kinds of things: - **Artifacts the model tends to add on its own.** Blur, compression noise, watermarks, stray text, extra limbs, six fingers. These are statistical habits, and naming them shifts the odds. - **Styles you are drifting toward by accident.** "Painting, illustration" when a photo comes out drawn; "photorealistic, 3d render" when an anime render goes glossy. - **Objects the subject attracts.** Glasses on a face you described as unadorned, a crowd behind a lone figure, a second cup on the table. It does not earn its keep on things the positive prompt already handles. If you want daylight, write "midday sun"; writing "night, darkness" in the negative is a weaker way to say the same thing. And a negative list that grows past a couple of lines starts fighting the positive prompt: the image gets flat, generic, and oddly cautious, because everything is a little bit forbidden. The rule: the positive prompt says what you want; the negative prompt says what the model keeps adding that you do not. ## What CFG scale does The [CFG scale](/glossary/cfg-scale) (classifier-free guidance) sets how strongly each step is pulled toward the prompt and, on the same dial, away from the negative. It is the "how literally" knob. - **Lower CFG** โ€” looser interpretation, softer contrast, more variety between takes, and a higher chance that a named detail goes missing. - **Higher CFG** โ€” more literal, more saturated, harder edges, and past a point, crunchy textures, doubled outlines and burnt highlights. Two practical consequences. First, negatives bite harder as CFG goes up, so if a negative term seems ignored, raising CFG one notch is often the fix rather than repeating the term. Second, photoreal work likes CFG a notch lower than illustration does: a photograph is allowed to be a little loose, while a drawing tolerates the saturation. Move it one notch at a time from the default. If a change of CFG also changes the composition, you moved too far; the dial is for obedience, not for layout. ## Where each control lives on eroq **Images.** All three image models, Image One (photoreal), Image Anime and Image Art, accept a negative prompt and a CFG value, alongside reference photos and a batch of 1 to 4 at 10 credits each. The [image API](/docs/images) exposes the same parameters, so a negative list that works in the studio works in your product. **Video.** Video engines are less uniform. The studio shows a negative prompt field only when the selected engine honors one; in the current roster, [Kling 2.5 Turbo](/models/kling-2-5-turbo) takes a native negative prompt (Creator plan and up). On the other engines, describe what is there instead of what is not: "an empty street" beats a negative "crowd" every time, on every engine. There is no CFG dial on video; the prompt structure does that job, and the [Video studio](/studio/video) folds camera and lighting picks into it for you. **Seeds.** Seedance engines honor a reproducibility [seed](/glossary/seed): the same prompt, settings and seed produce the same clip. That is what makes a fair comparison possible on video; change the negative prompt, keep the seed, and the difference you see is the negative prompt. [Seedance 2.5](/models/seedance-2-5) adds first and last frame control on top. Two things that look like negative-prompt problems and are not: a request outside the content policy returns `content_blocked` and is never charged, and a render that fails for any other reason refunds itself. Neither is fixed by the negative field. ## Practical lists Keep one base list per kind of work and add per-shot terms on top. These are the ones we reuse. **Photoreal (Image One)** > painting, illustration, cartoon, 3d render, plastic skin, waxy, oversaturated, extra fingers, deformed hands, blurry, text, watermark, logo **Anime and illustration (Image Anime)** > bad hands, extra fingers, missing fingers, extra limbs, deformed, bad anatomy, blurry, low quality, jpeg artifacts, text, watermark, signature, cropped, out of frame, photorealistic **Painterly (Image Art)** > photograph, 3d render, flat digital, oversharpened, text, watermark, extra limbs, blurry **Product shots** > hands, people, reflections of a camera, dust, fingerprints, text, watermark, busy background, extra objects **Video, on engines that accept a negative prompt** > text, watermark, logo, extra limbs, distorted face, split screen, jump cuts, flicker Per-shot additions are short and specific: "glasses", "beard", "second person", "rain". Add them for the shot and take them off afterward. ## A workflow that converges 1. Write the positive prompt, leave the negative empty, render a batch of four at the default CFG. 2. Look at what appeared that you did not ask for. Those words go in the negative prompt. 3. Look at what you asked for and did not get. Move that detail earlier in the prompt; if it is still missing, raise CFG one notch. 4. On video, lock the seed on a Seedance engine before comparing takes, so the only variable is the one you changed. 5. When a take is right, it is in your library with its full recipe (prompt, negative, CFG, seed); Remix reloads all of it for the next variation. Four renders per step, one change per step. It sounds slow and it is the fastest way there. ## FAQ ### Do negative prompts work on video? On engines that accept them, yes; the studio shows the field only for those engines, and Kling 2.5 Turbo is the one in the current roster that takes a native negative prompt. On the others, write what should be in the frame instead of what should not. ### Is a higher CFG always more accurate? No. It is more literal, which is not the same thing. High CFG oversaturates, hardens edges and eventually adds artifacts of its own. Raise it when a named detail keeps going missing; otherwise leave it alone. ### What is a seed and why does it matter here? A seed is the starting noise for a render. Same prompt, settings and seed on a Seedance engine produce the same result, so you can change a negative term and see exactly what it did instead of guessing across random takes. Open [the Image studio](/studio/image), paste the photoreal list into the negative field, and render a batch of four to see the difference.