Skip to content
ImagesFree

Six Compressed ComfyUI Prompt Rules From 200 Images

A practical guide distilling a image generator's experience after ruining 200 test images: weight syntax, the four-layer golden order, why prompts should be in English, a universal negative prompt template, anime and realistic style vocabularies, and quick fixes for common generation problems.

Stable Diffusion

Source previews

These example images come from the original source. DROPIDEA has not independently tested these examples; results depend on the model and settings.

Ready to personalize

Your prompt

English

Replace the red variables with your information. Copy the full prompt and use it with your preferred agent.

TXT MD English version
After testing and discarding 200 images, Long-ge compressed ComfyUI prompts into 6 rules. At first I thought prompts were just stacking keywords—"二次元、动漫、可爱、JK制服、长发"—dumping them all in, and the output was always just a little off from what I wanted. Later I figured it out: the model doesn't understand natural language, it only recognizes keywords + probability weights. ❶ Weight syntax, one line is enough (keyword:number)—1.2 to strengthen, 0.8 to weaken, 1.0 is default. Nesting no more than 3 levels (deeper causes image overfitting). Minimalist shorthand: (keyword++) = 1.2, [keyword] = 0.9. ❷ Four-layer golden order Subject → Environment → Style → Quality. The model pays the most attention to the first 20 tokens; core words must go first. 1girl, long pink hair, smiling, kimono, cherry blossom garden, soft sunset light, anime style, cel shading, pastel palette, masterpiece, best quality, 4k, intricate details ❸ Don't use Chinese prompts SD1.5/SDXL training data is mainly in English; Chinese tag recognition rate is extremely low. English prompt + Chinese annotations is the optimal solution. ❹ Universal negative prompt template (enough for 90% of scenarios) low quality, bad anatomy, extra fingers, deformed eyes, blurry, ugly, duplicate, watermark, signature Anime additions: nsfw, monochrome, flat colors Realistic additions: plastic skin, wax face, overexposed, noise ❺ Two style vocabulary sets - Anime: anime screencap, cel shading, vibrant colors, detailed background - Realistic: photorealistic, cinematic lighting, 85mm, subsurface scattering, film grain ❻ Image problem troubleshooting (lifesaving) - Irrelevant → raise CFG to 7-8 - Single element overcooked → lower core word weight / move it back - Multi-character merging → rewrite the four-layer structure - Repeated objects → change seed + add "repeat" to negatives 6 rules isn't a lot, copy and paste and you can use them. Whether it's worth it or not, I can't say—but after I ruined 200 images I can stably produce good output.
Preview results

Your prompt

After testing and discarding 200 images, Long-ge compressed ComfyUI prompts into 6 rules. At first I thought prompts were just stacking keywords—"二次元、动漫、可爱、JK制服、长发"—dumping them all in, and the output was always just a little off from what I wanted. Later I figured it out: the model doesn't understand natural language, it only recognizes keywords + probability weights. ❶ Weight syntax, one line is enough (keyword:number)—1.2 to strengthen, 0.8 to weaken, 1.0 is default. Nesting no more than 3 levels (deeper causes image overfitting). Minimalist shorthand: (keyword++) = 1.2, [keyword] = 0.9. ❷ Four-layer golden order Subject → Environment → Style → Quality. The model pays the most attention to the first 20 tokens; core words must go first. 1girl, long pink hair, smiling, kimono, cherry blossom garden, soft sunset light, anime style, cel shading, pastel palette, masterpiece, best quality, 4k, intricate details ❸ Don't use Chinese prompts SD1.5/SDXL training data is mainly in English; Chinese tag recognition rate is extremely low. English prompt + Chinese annotations is the optimal solution. ❹ Universal negative prompt template (enough for 90% of scenarios) low quality, bad anatomy, extra fingers, deformed eyes, blurry, ugly, duplicate, watermark, signature Anime additions: nsfw, monochrome, flat colors Realistic additions: plastic skin, wax face, overexposed, noise ❺ Two style vocabulary sets - Anime: anime screencap, cel shading, vibrant colors, detailed background - Realistic: photorealistic, cinematic lighting, 85mm, subsurface scattering, film grain ❻ Image problem troubleshooting (lifesaving) - Irrelevant → raise CFG to 7-8 - Single element overcooked → lower core word weight / move it back - Multi-character merging → rewrite the four-layer structure - Repeated objects → change seed + add "repeat" to negatives 6 rules isn't a lot, copy and paste and you can use them. Whether it's worth it or not, I can't say—but after I ruined 200 images I can stably produce good output.

What you get

  • A simple explanation of weight syntax including minimalist shorthand like (keyword++) and [keyword].
  • A complete example prompt demonstrating the four-layer golden order.
  • A copy-ready universal negative prompt template covering 90% of scenarios with anime and realistic add-ons.
  • Two separate style vocabularies for anime and photorealistic work.
  • A quick diagnostic list for four common generation problems with direct fixes.

Use cases

  • Improving ComfyUI prompt quality for SD1.5 and SDXL users.
  • Building a personal library of negative and style keywords.
  • Training beginners on structured prompting instead of random keyword stacking.
  • Quickly diagnosing generation issues such as overcooked elements or repeated objects.

How to use

  • Read the six rules in order to grasp the fundamentals of building ComfyUI prompts.
  • Memorize the weight syntax (keyword:number) and its limits: 1.2 to strengthen, 0.8 to weaken, with no more than 3 nesting levels.
  • Apply the four-layer order to your prompts: subject, then environment, then style, then quality, keeping core words within the first 20 tokens.
  • Copy the universal negative prompt template and add the anime or realistic additions depending on your style.
  • Use the two style vocabulary sets as ready-made sources for describing anime or photorealistic output.
  • Refer to the troubleshooting section when results are irrelevant, an element is overcooked, characters merge, or objects repeat.

Related prompts