Chinese AI-art prompt weight editor: set (term:1.3) emphasis or bracket de-emphasis for Chinese and English terms. For Stable Diffusion WebUI and ComfyUI.
Chinese AI Art Prompt Weight Editor
Wrap Chinese prompt terms with weight syntax for AI-art models. Type comma-separated terms, drag a slider per term, and the tool assembles a weighted prompt — (词:1.3) to emphasise, (词:0.7) to de-emphasise — ready to paste into Stable Diffusion WebUI, Forge or ComfyUI. Everything is built in your browser; nothing is sent to a server and no model is called.
Tip: this editor only assembles text. Copy the result into Stable Diffusion WebUI / Forge / ComfyUI yourself — no image is generated, no model is called, and nothing is sent anywhere.
Type some comma-separated terms above — a weight slider will appear for each one.
Setting a weight per term, then picking the syntax
Type comma-separated terms
Write your Chinese art terms in the text box, separated by a Chinese comma "，" or an ASCII comma ",", e.g. "夜晚的城市，霓虹灯，雨，电影感，细节丰富". Each term becomes its own row with a weight slider, and duplicate terms are removed automatically.
Set a weight per term with the slider
Drag a slider right (above 1.0) to emphasise a term and left (below 1.0) to de-emphasise it; 1.00 means no weighting. A common emphasis range is 1.1–1.5 — going too high (above ~1.8) tends to unbalance the image. Each row shows its current value live, and Reset returns it to 1.00.
Choose the weight syntax style
Pick a style: Explicit weights output forms like (词:1.3), supported directly by Stable Diffusion WebUI (AUTOMATIC1111), Forge and ComfyUI; Bracket style uses (term) to emphasise and [term] to de-emphasise, but only in steps of ×1.1; AUTOMATIC1111 and Forge read both, while ComfyUI reads only the round brackets and treats [term] as plain text. You can switch styles at any time without re-typing.
Copy into your art model
Click Copy and paste the weighted prompt into the positive-prompt box of Stable Diffusion WebUI, Forge or ComfyUI. NovelAI writes weights differently ({term} to emphasise, [term] to weaken, ×1.05 a layer), and hosted models such as Alibaba’s Wan (通义万相) document no weight syntax. Everything is assembled locally in your browser; nothing is sent to any server and no model is called.
How the Chinese prompt weight editor works
Weighting tells the model what matters most
In AI-art tools such as Stable Diffusion WebUI and ComfyUI, a prompt is a list of terms — but not every term should pull with equal force. Weighting is how you tell the model which words matter more and which should fade into the background. The explicit form (词:1.3) sets a term's weight to 1.3, a little more emphasis than the default of 1.0; (词:0.7) pulls it down. The bracket form does the same thing through nesting: (词) adds roughly ten percent of weight per pair of parentheses, while [词] removes roughly ten percent per pair of square brackets. This editor keeps that bookkeeping out of your way: you type your terms once, separated by commas, and each one gets its own slider. Drag the slider, watch the number update, and the weighted prompt rebuilds itself instantly, in whichever syntax style your target model expects.
Chinese terms weight in exactly the same way as English ones, because the model only ever sees bracketed plain text — (霓虹灯:1.3) is parsed no differently from (neon lights:1.3). What changes is how well a given model understands the underlying Chinese vocabulary, which is why prompting a Chinese-native model in Chinese, with weights, often gives the most faithful result. The editor also removes duplicate terms automatically, so you never accidentally weight the same word twice, and it accepts both the Chinese comma "，" and the ASCII comma "," as separators. Because the output is ordinary text, the same weighted prompt is portable: paste it into the positive-prompt box of almost any modern image model and it will be understood.
"A weight is not a volume knob to crank — it is a way to say, of all these words, this one matters a little more. Used sparingly, it transforms an image; used everywhere, it does nothing at all."
Restraint and balance beat a single huge weight
The mistake almost everyone makes early on is treating weight as a volume knob and turning it to the maximum. Push a single term to 1.8 or beyond and it stops being emphasis and starts being domination: the weighted element crowds out everything else, the composition collapses, and artefacts creep in. A far more reliable habit is restraint. Pick the one or two terms that genuinely carry the image, lift them gently into the 1.1–1.5 range, and leave everything else at 1.0. Weight is relative — if you raise every term at once, you have emphasised nothing, you have simply rescaled the whole prompt. The slider in this editor is deliberately capped at a sane maximum and steps in small increments precisely to encourage that fine-tuning mindset over brute force.
Weighting also works best in partnership with the other levers you already have. Term order matters: words near the front of a prompt tend to receive more attention, so a key term placed early and given a modest 1.2 often beats the same term buried at the end at 1.6. De-emphasis is the quiet half of the craft — dropping an unwanted-but-stubborn element to 0.6, or wrapping it in square brackets, is frequently cleaner than deleting it outright and hoping it stays gone. Treat your first weighted prompt as a draft: generate, see where the image drifts, then nudge one weight up or down and try again. Two or three rounds of that, with everything assembled instantly and privately in your browser, will usually land you on a balanced, reusable prompt that says exactly what you meant.
Weight is relative — raise everything and you raise nothing
Weights let you tell an art model which terms matter more: (词:1.3) amplifies a term and (词:0.7) plays it down — without rewriting the whole prompt.
The explicit form (term:value) is the most portable weighting syntax across Stable Diffusion, ComfyUI and Forge, with values usually between 0.5 and 1.5.
Bracket style expresses strength by nesting: (term) ≈ ×1.1, ((term)) ≈ ×1.21, and [term] ≈ ÷1.1 — AUTOMATIC1111's rule, which Forge shares; ComfyUI reads only the round brackets, not [term]. NovelAI instead uses {term}, ×1.05 a layer.
Higher is not always better: pushing one term above 1.8 often crowds out other elements, produces artefacts or breaks the composition; 1.1–1.4 is usually enough.
Chinese terms weight exactly like English ones — the model only reads bracketed plain text, so (霓虹灯:1.3) behaves just like (neon lights:1.3).
De-emphasis (below 1.0, or square brackets) is ideal for suppressing things you do not want but cannot fully delete: clutter, an over-strong colour, a stylistic tic.
Weight is relative: nudging many terms to 1.4 at once emphasises none of them — the effective move is to highlight just one or two key words.
Term order matters too: earlier terms tend to get more attention, so combining weight with order is more reliable than relying on either alone.
Models tolerate extreme out-of-range weights differently; start with gentle values and make small adjustments, closing in on the image you want.
This tool assembles the prompt entirely in your browser — your input is never uploaded, never sent to a model, and never stored.
Frequently Asked Questions
- No. It only joins the Chinese terms you type into a prompt carrying the weights you set, entirely in your browser. It does not generate images and does not call Stable Diffusion, NovelAI or any model. You copy the generated prompt and use it in your own art tool.
- (词:1.3) sets the term's weight to 1.3 — more emphasis than the default; a value below 1 (e.g. 0.7) de-emphasises it. [词] is the bracket-style way to de-emphasise, with each extra square bracket reducing the term's influence by roughly a tenth; the matching (词) emphasises. You choose which to output under "syntax style".
- For emphasis, 1.1 to 1.5 is common; for de-emphasis, 0.5 to 0.9. The more extreme the weight, the more it disturbs the image. Usually highlighting just one or two key terms while leaving the rest at 1.0 is the most stable. Exact values vary by model, so adjust in small steps.
- Explicit weights (term:value) work with Stable Diffusion, ComfyUI, Forge and Automatic1111; bracket-style (term) works in all of them, but [term] de-emphasis only in AUTOMATIC1111 and Forge — ComfyUI reads it as plain text. NovelAI writes weights differently ({term}, ×1.05 a layer), and hosted models such as Wan (通义万相) document no weight syntax. Both are just plain text and are language-neutral, so Chinese terms work the same way.
- Syntactically they are identical — the model only reads bracketed plain text. The real effect depends on how well the model understands the Chinese words: models with native Chinese support respond better to Chinese weighting, while English-only models may need you to switch to English terms.
- An over-high weight (above ~1.8) lets one term dominate the whole image, crowding out other elements, creating artefacts or breaking composition. Drop back to 1.1–1.4 and balance with term order and de-emphasis, rather than just raising a single number.
- No. All splitting, weighting and assembly happen locally in your browser with plain JavaScript. The terms and weights you enter are never sent to any model, server or third party, and nothing is stored.
- The tool removes duplicates automatically, keeping the first occurrence's position so a term is not accidentally amplified by being weighted twice. If you genuinely want stronger emphasis, raise that term's weight rather than repeating the word.
- Yes. Both the Chinese comma "，" and the ASCII comma "," separate terms and are treated identically. Spaces around each term are trimmed automatically and blank terms are ignored.
- Completely free, with no account or sign-up and no usage limit. It runs in your browser and collects no data.
Method & sources
How it computes
Terms split on ASCII and full-width commas and de-duplicated; a term pasted already weighted ((猫:1.2), ((猫)), [猫]) is unwrapped to its weight. Explicit style writes (term:w); bracket style writes n pairs of () or [] with n = round(|log w / log 1.1|), at most 5, and shows the weight that gives. Literal brackets are escaped. AUTOMATIC1111 and Forge read both styles; ComfyUI reads () only (PromptWeightKernel). No model is called.
What this tool implements
- Corrected 2026-10-25 (code): bracket mode silently capped at 5 pairs and always added at least one, so a slider at 2.00 gave 1.61 and 1.02 gave 1.10 with no sign of it; the row now shows the weight the brackets actually give, and near-1 weights stay unbracketed. Literal brackets in a term were not escaped ('anime (character)' became two weighted terms); a pasted '(猫:1.2)' was weighted again as '((猫:1.2):1.3)'. All outputs checked with AUTOMATIC1111's own prompt parser.
- Corrected 2026-10-25 in all three languages and the SEO descriptions: the page said bracket style suits NovelAI and that the syntax works in 'most' Chinese text-to-image models, Jimeng and Wan. NovelAI uses {} and [] at ×1.05 per layer; Alibaba's Wan documentation describes no weight syntax and rewrites prompts by default; Jimeng's documentation could not be read. The page now names AUTOMATIC1111, Forge and ComfyUI.
- The weight ranges the page suggests (1.1-1.5, trouble above about 1.8) are practitioners' guidance, not a vendor rule.
- Corrected 2026-10-26 in all three languages: the page said [词] de-emphasis works in ComfyUI; ComfyUI's README documents () only and its parser has no square-bracket rule, so [词] is plain text there. The intro now gives the explicit de-emphasis form (词:0.7), which every named tool reads.
Sources
- AUTOMATIC1111. stable-diffusion-webui, modules/prompt_parser.py, parse_prompt_attention: '(abc) - increases attention to abc by a multiplier of 1.1'; '[abc] - decreases attention to abc by a multiplier of 1.1'; '\( - literal character '('', read 2026-10-26: https://raw.githubusercontent.com/AUTOMATIC1111/stable-diffusion-…
- AUTOMATIC1111. stable-diffusion-webui wiki, Features, Attention/emphasis: 'a ((word)) - increase attention to word by a factor of 1.21 (= 1.1 * 1.1)'; 'Specifying weight only works with () not with []'; 'use the backslash to escape them', read 2026-10-26: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Feat…
- ComfyUI README: 'You can use () to change emphasis of a word or phrase like: (good code:1.2) or (bad code:0.8). The default emphasis for () is 1.1. To use () characters in your actual prompt escape them', read 2026-10-26: https://github.com/Comfy-Org/ComfyUI
- ComfyUI. comfy/sd1_clip.py, parse_parentheses / token_weights: round brackets only (×1.1, or the weight after the last colon); square brackets have no rule and stay text, read 2026-10-26: https://github.com/Comfy-Org/ComfyUI/blob/master/comfy/sd1_clip.py
- NovelAI documentation. Strengthening and weakening: 'multiplied by 1.05 if you enclose the tags or text … with { and }'. https://docs.novelai.net/en/image/strengthening-weakening/
- Hugging Face diffusers. Prompt weighting: weights need a helper library, and 'Prompt weighting does not always help' on Flux-class models. https://huggingface.co/docs/diffusers/using-diffusers/weighted_pr…
- Alibaba Cloud Model Studio. Wan text-to-image API reference: prompt_extend defaults to true (智能改写); no weight syntax documented. https://help.aliyun.com/zh/model-studio/text-to-image-v2-api-refe…
Pick up where you left off
Stored only in this browser — never sent to our servers.