Skip to main content
UFOZoo

Stable Diffusion Prompt Generator: 48 Tags & Weights

Stable Diffusion prompt generator online: build prompts with subject, style, quality, lighting, and camera tags, with weights and negative prompts.

Updated 2026-08-16

Related Tools

AI Video Prompt Generator: Prompt Builder for Text-to-Video

System Prompt Generator: Build Role & Rules Prompts

Prompt Sanitizer: Remove Sensitive Data from AI Prompts

Claude Prompt Generator: System Prompts & XML Examples

Midjourney Prompt Generator: Parameters & Styles Online

AI Prompt Compressor: Reduce Tokens & Save on LLM Costs

Features

  • Categorized tag library: 48 curated SD tags across Subject, Style, Quality, Lighting, Camera, and Color, each with a Chinese display name
  • Weight syntax: set 0.8 / 1.0 / 1.2 / 1.5 per tag; non-default weights are emitted as (word:1.2) syntax, supported by AUTOMATIC1111 and ComfyUI
  • Negative prompt builder: a built-in blacklist (bad quality, worst quality, blurry, deformed, extra limbs...) plus your own custom negative tags
  • Shuffle order: randomize tag order with one click; order matters in SD because the model attends to earlier tags more strongly
  • Custom tags: type your own comma-separated tags and they are merged into the positive prompt automatically
  • ComfyUI compatible: the output is plain comma-separated tags with standard weight syntax, usable in AUTOMATIC1111, ComfyUI, and NovelAI-style tools
  • Chinese UI: the interface and tag names are localized; tags themselves stay in English because the SD ecosystem trains on English tags (honest note included)
  • Separate copy or copy all: copy the positive prompt, the negative prompt, or both together, and download the pair as a text file
  • Live output: the prompt rebuilds instantly as you toggle tags and weights, with an empty-state guard when nothing is selected
  • Works offline after first load

How to Use

  1. 1Pick tags on the left panel: open a category (Subject, Style, Quality, Lighting, Camera, Color) and click the chips you want; each selection shows its English SD tag.
  2. 2Fine-tune weights: for each selected tag choose 0.8 to weaken or 1.2 / 1.5 to emphasize; weight 1.0 keeps the tag plain without syntax.
  3. 3Add your own tags in the Custom tags box, comma-separated; they are appended to the positive prompt in the order you type them.
  4. 4Optionally extend the negative prompt with your own comma-separated tags; the default blacklist is always included.
  5. 5Click Shuffle order to randomize the tag sequence; in SD the position of a tag influences how strongly the model attends to it.
  6. 6Copy the positive prompt and paste it into the text prompt box of AUTOMATIC1111, ComfyUI, or your SD WebUI.
  7. 7Copy the negative prompt into the negative prompt box of your SD tool to suppress unwanted content like blur and extra limbs.
  8. 8Use Copy all or Download to grab the full prompt pair for saving or sharing.
  9. 9Making a realistic portrait: pick 1girl + photorealistic + portrait shot + studio lighting + golden hour, set close-up to 1.2, then add your own details like 'red dress'.
  10. 10Making an anime illustration: pick 1boy + anime style + masterpiece + soft lighting + vibrant colors, and set anime style to 1.5 for a stronger style lock.
  11. 11Building a negative-prompt blacklist: start from the built-in defaults, add your project-specific exclusions (e.g., 'text, logo'), and reuse the same negative text across all your generations.
  12. 12Reset clears everything, including selections, weights, custom tags, and custom negative tags, so you can start a new prompt from scratch.

Frequently Asked Questions

How does the weight syntax work?

A weight wrapped in parentheses changes how strongly the model attends to that tag: (word:1.2) strengthens it and (word:0.8) weakens it, while a plain word without parentheses is the default weight of 1.0. Both AUTOMATIC1111 and ComfyUI support this syntax. In this tool, choosing 0.8, 1.2, or 1.5 for a tag adds the parentheses automatically; weight 1.0 leaves the tag plain.

Does tag order really matter?

Yes, but the effect is subtle. Stable Diffusion models process the prompt as a sequence of tokens and tend to attend more strongly to tags near the beginning. Putting your most important tags first usually improves adherence, which is why this tool offers a Shuffle order button. If a tag keeps getting ignored, move it earlier or raise its weight instead.

Why are the tags in English even though the UI is Chinese?

Stable Diffusion models were trained on massive datasets whose labels are predominantly English (Danbooru tags, LAION captions, and similar). English tags therefore produce noticeably more reliable results than Chinese ones. The interface and tag names are fully localized to Chinese, but the tags themselves stay in English by convention; that is the honest trade-off of the SD ecosystem.

What is the difference between this tool and the Midjourney prompt generator?

The two tools produce prompts for different engines with different syntax. Midjourney uses parameters such as --ar 16:9, --v 6, and --style raw attached to a natural-language description. Stable Diffusion uses comma-separated English tags with optional (word:1.2) weights and a separate negative prompt. This tool builds the tag-and-weight format that AUTOMATIC1111 and ComfyUI expect.

What is a negative prompt for?

A negative prompt lists what you do not want in the image, and the model actively avoids those concepts during generation. The built-in blacklist covers common failures such as bad quality, worst quality, blurry, deformed, extra limbs, and watermarks. Add your own exclusions (for example 'text' or 'logo') and reuse the same negative prompt across generations for consistency.

Does this work with ComfyUI?

Yes. The output is plain comma-separated tags with standard (word:weight) syntax, which is exactly what the CLIP text encoder in ComfyUI understands. Paste the positive prompt into a CLIP Text Encode node and the negative prompt into the negative input of the same node.

Can I use Chinese tags in the prompt?

Technically yes, the model will process any text you give it, but results are usually worse than with English tags. Because the training data is overwhelmingly English-labeled, Chinese tags are interpreted less precisely. This tool deliberately keeps tags in English for quality, while the interface and tag names are in Chinese so you can understand and choose them.

Why do some tags not seem to affect the image?

Several reasons: the tag may be out of the model's vocabulary, the weight may be too low, or the tag sits late in a very long prompt where attention is diluted. Try moving the tag to the front, raising its weight to 1.2 or 1.5, or removing unrelated tags so the model can focus. With SD, less clutter usually means better adherence.

What weights should I use?

Start with plain tags (weight 1.0) and only adjust when something is off. Use 1.2 for a tag the model keeps ignoring, 1.5 for the single most important element such as the art style, and 0.8 to soften a tag that overpowers the image. Weights above 1.5 tend to cause artifacts, so this tool caps the range at 1.5.

Can I generate prompts from an image?

No. This tool builds prompts from its tag library and does not analyze images. If you want tags for an existing image, use the Interrogate feature in AUTOMATIC1111 (CLIP or DeepDanbooru mode) or a dedicated CLIP interrogator, then paste the resulting tags into this tool to tune weights and build the negative prompt.

Is there a Chrome extension version of this prompt generator?

No, and you do not need one: this is a plain web page that runs entirely in your browser and keeps working offline after loading, so a bookmark gives you the same one-click access without installing anything. Browser extensions that add tags directly inside AUTOMATIC1111 exist, but this tool's output can be pasted into any WebUI regardless.

Can I use the generated prompt in other image AI tools?

The tag-and-weight format is understood by AUTOMATIC1111, ComfyUI, and most Stable Diffusion WebUIs. Other tools with different syntaxes (Midjourney parameters, DALL-E natural language) will accept the text but interpret it differently; for those, a plain natural-language prompt usually works better.

Why does this stable diffusion prompt generator free tool produce different images even when I reuse the same prompt?

The generator only writes the prompt. The image is produced by the Stable Diffusion model, whose sampler starts from random noise unless you fix the seed. Same prompt + different seed (or different sampler, CFG, or model) = different image. To reproduce an exact image, copy the prompt AND the seed, sampler, CFG scale, and model name from your WebUI. If a tag seems ignored, check the negative prompt and whether the word is in the model's vocabulary.