Free Stable Diffusion Prompt Generator (With Negatives)

TL;DR: This free Stable Diffusion prompt generator is a prompt, not a website. Paste it into ChatGPT, Claude or Gemini, describe your image, and name your setup. It returns a positive prompt, a matched negative prompt and starting settings, formatted for SD 3.5 on Stability's API, SDXL, or SD 1.5 in AUTOMATIC1111 or ComfyUI.

Most Stable Diffusion prompt generator pages give you a box, a button and a pile of adjectives. Then you paste the result into your setup and half of it does nothing, because the tool had no idea whether you were calling Stability's hosted API, running SDXL in ComfyUI, or still on SD 1.5 with a folder of LoRAs.

This page takes a different approach. Below is a generator you run inside the chatbot you already use. It asks what you are generating with, writes the prompt in that setup's format, and pairs it with a negative prompt aimed at the failures that setup actually produces. After the generator come filled examples, a negative prompt bank sorted by problem, and the handful of technical rules that decide whether a negative prompt does anything at all.

Every technical claim here was checked on 17 September 2026 against Stability AI's API specification, the Hugging Face diffusers documentation and the AUTOMATIC1111 wiki.

What does this Stable Diffusion prompt generator actually produce?

Four things, every time:

1. A positive prompt in the style your model reads best. Descriptive sentences for SD 3.5, comma-separated phrases for SDXL and SD 1.5.
2. A negative prompt built from the failures most likely for your subject, ordered most important first and kept short.
3. Starting settings: aspect ratio, guidance and, for local setups, a sampler and step range to begin testing from.
4. Two variations that change one thing each, so you can compare instead of guessing.

It does not generate the image. You take the output to Stability's API, ComfyUI, AUTOMATIC1111 or whatever you run. That separation is the point. Writing the prompt and rendering the image are different jobs, and a generator that knows your renderer writes a better prompt.

The generator: copy and paste this

Copy the whole block into a new chat. Replace the three bracketed lines at the bottom, then send it.

You are a Stable Diffusion prompt engineer. Write prompts that a specific
Stable Diffusion setup will read correctly.

STEP 1 - Identify the setup from SETUP below. Use these rules:

A) "SD 3.5 API" (Stability AI hosted: Ultra, Core or SD3 endpoints)
   - Positive prompt: 1 to 3 descriptive sentences. Subject first, then
     setting, lighting, camera or medium, mood, colour palette.
   - Weights: only if asked. Stability's Core endpoint uses (word:weight)
     with weight between 0 and 1. Never use weights above 1 here.
   - Negative prompt: 5 to 15 comma-separated terms, most important first.
   - Settings: suggest aspect_ratio and, for the SD3 endpoint, cfg_scale.
     Warn that Turbo and Flash models default to cfg_scale 1, where a
     negative prompt may have little or no effect.

B) "SD 3.5 local" (diffusers or ComfyUI)
   - Same positive style as A. Keep the key subject and style words in the
     first sentence, because the CLIP encoders truncate at 77 tokens even
     though T5 reads further.
   - Negative prompt: 5 to 15 terms. Remind me that negatives are skipped
     unless guidance is set above 1.

C) "SDXL" or D) "SD 1.5" (AUTOMATIC1111, Forge, ComfyUI)
   - Positive prompt: comma-separated phrases, most important first.
     Subject, details, setting, lighting, style, quality terms.
   - Weights allowed in (word:1.2) form, between 0.5 and 1.5. Use at most
     three weighted terms.
   - Negative prompt: 8 to 20 terms, most important first.
   - Settings: resolution suited to the model, a sampler and step range to
     start testing from, and a guidance value.

STEP 2 - Build the negative prompt from the failures this subject usually
produces (hands, faces, text, extra objects, lighting, framing). Name the
specific failure, not a vague quality word. No sentences. Never use "no".

STEP 3 - Output exactly this format:

POSITIVE:
<prompt>

NEGATIVE:
<prompt>

SETTINGS:
<bullet list>

VARIATIONS:
1. <same prompt with one change, and what the change tests>
2. <same prompt with a different single change>

WATCH FOR:
<one sentence on the most likely failure for this subject>

SETUP: [SD 3.5 API / SD 3.5 local / SDXL / SD 1.5]
IMAGE: [describe the image in plain words]
MUST AVOID: [anything you already know goes wrong, or "nothing yet"]

The generator deliberately asks for your setup first. The same scene should be written differently for SD 3.5 and SDXL, and the weight syntax that works in one setup is outside the documented range in another. More on that below.

What does the output look like for real subjects?

Here are three filled runs, lightly trimmed, so you can see the shape before you try it. Treat them as starting points. Seeds, samplers and checkpoints change results more than any single word.

SD 3.5 on Stability's API, product shot:

POSITIVE:
A matte black ceramic coffee mug on a pale oak table beside a folded linen
napkin, soft morning window light from the left, shallow depth of field,
clean minimalist product photograph, neutral warm palette.

NEGATIVE:
text, logo, watermark, cluttered background, harsh shadows, reflections on
mug, distorted handle, oversaturated

SETTINGS:
- Endpoint: core, aspect_ratio 4:5
- If you switch to the sd3 endpoint with sd3.5-large, try cfg_scale 4 to 6
- On Turbo or Flash, test the negative prompt on and off before relying on it

SDXL in ComfyUI, fantasy portrait:

POSITIVE:
portrait of an elderly elven cartographer, (weathered face:1.2), silver
braided hair, holding a brass compass, candlelit study, maps pinned to wall,
warm rim light, oil painting style, detailed brushwork

NEGATIVE:
fused fingers, extra fingers, deformed hands, blurry face, cross-eyed,
modern clothing, text, signature, watermark, flat lighting, lowres

SETTINGS:
- 832x1216
- Start at 25 to 30 steps with a DPM++ 2M family sampler
- Guidance 5 to 7

SD 1.5, anime landscape:

POSITIVE:
anime style, wide shot of a mountain village at dusk, paper lanterns,
cherry blossom petals in the air, river in foreground, (soft pastel sky:1.1),
detailed background, cel shading

NEGATIVE:
people, characters, text, speech bubble, blurry, jpeg artifacts,
oversaturated, dark, lowres, bad perspective

SETTINGS:
- 768x512
- 20 to 30 steps, Euler a to start
- Guidance 7

If you want to understand those step and sampler suggestions rather than copy them, read the guides on what sampling steps actually do and the Stable Diffusion sampler reference. One number to note from the first: Stability's hosted image endpoints do not expose a steps parameter at all, which is why the SD 3.5 API example has none.

Which negative prompts fix which problems?

Pick by the failure you are seeing, not by habit. Each line below is short on purpose, so you can stack two or three without drowning the important terms.

HANDS AND ANATOMY
fused fingers, extra fingers, missing fingers, deformed hands, extra limbs

FACES
asymmetrical eyes, cross-eyed, distorted face, blurry face, plastic skin

TEXT AND BRANDING
text, letters, watermark, signature, logo, caption

COMPOSITION
cropped head, cut off, out of frame, tilted horizon, cluttered background

EXTRA OBJECTS
people, crowd, animals, cars, furniture   (keep only what applies)

IMAGE QUALITY
blurry, lowres, jpeg artifacts, noise, pixelated

LIGHTING AND COLOUR
flat lighting, harsh shadows, overexposed, oversaturated, washed out

STYLE DRIFT (photo you wanted, illustration you got)
illustration, painting, cartoon, 3d render, cgi

STYLE DRIFT (illustration you wanted, photo you got)
photograph, photorealistic, realistic skin texture, film grain

Two rules make this list work. First, name the concrete thing. A term like bad quality carries almost no visual information, while jpeg artifacts does. Second, subtract terms that do not apply. If there are no people in your scene, extra fingers is noise.

A third rule applies to length. Stability's hosted endpoints accept a negative prompt of up to 10,000 characters and label the field "an advanced feature", so length is not the constraint there. The constraint is attention: every extra term competes with the ones you actually care about. On local SDXL and SD 1.5 setups the encoders are CLIP-based, so put the two or three terms that fix your real problem at the very front, and treat everything after them as optional.

For a larger, cross-model library, including rewrites for models that have no negative field at all, see the free negative prompt generator for image models. If a term keeps being ignored, work through why negative prompts get ignored.

When does a Stable Diffusion negative prompt actually work?

Only when guidance is on. This is the rule most generator pages skip, and it explains a lot of frustration.

A negative prompt works through classifier-free guidance: the sampler steers away from what the negative describes. The diffusers documentation for both the SDXL and SD3 pipelines says the negative prompt is "Ignored when not using guidance". The parameter text describes that as a guidance scale below 1, but the SD3 pipeline's source code is stricter: it only applies guidance when the scale is greater than 1. So at exactly 1, a local negative prompt is skipped entirely.

Now look at the fast models. Stability's API specification says the Turbo and Flash models on its SD3 endpoint use a default cfg_scale of 1, while Large and Medium default to 4. At that low guidance setting, the documentation gives you no reason to expect a negative prompt to do much. I have not found a Stability page that states the effect either way, so treat it as something to test: render the same seed with and without the negative and compare.

The practical rules:

• Large and Medium models: negatives are worth writing. Keep guidance at or near the default.
• Turbo and Flash models: do not rely on the negative. Put the important exclusions into the positive prompt as positive descriptions instead. Plain white background beats cluttered background in the negative.
• Local setups: if you have lowered guidance to 1 for speed, your negative prompt has probably stopped doing anything.

The negative prompt support matrix covers how other image models handle this, including several that publish no negative field at all.

How should prompts differ between SD 3.5, SDXL and SD 1.5?

The models read text differently, so the same words are not equally effective.

Sources: Stability API specification, diffusers SD3 and SDXL pipeline docs, AUTOMATIC1111 wiki. Checked 17 September 2026.

SD 3.5 reads longer, more natural descriptions. Its pipeline adds a T5 encoder, which diffusers says reads 256 tokens by default and can be raised. The catch is in the same documentation: "The prompt with the CLIP Text Encoders is still truncated to the 77 token limit." So even on SD 3.5, put the subject and style in the first sentence, where every encoder sees them. For a fuller treatment of the newest model family, see Stable Diffusion 3.5 prompting in 2026.

SDXL has two encoders and lets you address them separately. Diffusers accepts a second prompt and a second negative prompt for the second encoder. Most people never use this. If you do, a common split is subject and content in the first, style and medium in the second.

Weights are not portable. The AUTOMATIC1111 wiki documents parentheses as a 1.1 multiplier and explicit values such as 1.5 to strengthen or 0.25 to weaken a term. Stability's Core endpoint documents the same bracket shape with a different range: the weight "is a value between 0 and 1", used to balance terms, as in its example of a sky that is more green than blue. A prompt with (sunset:1.4) written for AUTOMATIC1111 is outside that documented range on Stability's API. The generator above writes weights only after it knows your setup for exactly this reason.

How do you improve a generated prompt instead of rerolling?

Change one thing at a time and keep the seed fixed. Rerolling with new seeds tells you about luck, not about your prompt.

A simple loop:

1. Fix the seed and render the generator's output.
2. Name the single biggest problem in the image, in two or three words.
3. If the problem is something present that should not be, add it to the front of the negative prompt.
4. If the problem is something missing, describe it in the positive prompt. Do not try to fix absence with the negative.
5. Render again with the same seed. Keep the change only if it helped.

Then ask the chatbot for help with the diagnosis. After the first render, paste this follow-up into the same conversation:

I rendered your prompt. The result has this problem: [describe it].
Change the smallest possible part of the POSITIVE or NEGATIVE prompt to fix
it. Show only the changed lines and explain in one sentence why the change
should help. Do not change the seed or settings.

This keeps the conversation focused on one variable. It also leaves you with a record of which edits worked, which is the beginning of a personal prompt library. For wording that controls framing and camera position, the list of composition and framing terms is a useful companion.

How do you use the generator for a batch of images?

Keep the rules block fixed and change only the last three lines. The generator is written so that the instructions at the top never need editing, which makes it easy to run a series.

For a set of product images, for example, send the full generator once with the first product. For every product after that, send only a short follow-up in the same chat:

Same SETUP and rules. New IMAGE: [next product, same style as before].
MUST AVOID: [same list, plus anything the last render got wrong].
Keep the lighting, palette and camera wording identical to the previous
POSITIVE so the set looks consistent.

That last instruction is the one that matters for a batch. Consistency across a set comes from repeating the same style and lighting phrases word for word, not from asking the model to be consistent. If the chatbot starts paraphrasing, paste the previous positive prompt back in and tell it to change only the subject.

Two more habits help with batches. Keep one negative prompt for the whole set unless a specific image shows a specific failure, because changing the negative per image makes the set drift. And note the seed of any render you like: with the same seed, settings and style words, the next subject tends to inherit the look you approved, which is far cheaper than rediscovering it.

What should you tell the generator when you already know what goes wrong?

Use the MUST AVOID line, and be specific. It is the part of the generator most people leave as nothing yet, and it is where most of the value is once you have rendered a few images.

Good entries describe a visible failure: the model keeps adding a second cup, faces are too smooth, the logo area fills with fake text. The generator turns each into a short negative term and, where a negative is unlikely to work, a positive description to add instead. A second cup becomes extra cup in the negative and a single mug, alone on the table in the positive. Doing both is more reliable than either alone, and it keeps working if you later switch to a Turbo model where the negative carries less weight.

Poor entries are vague quality complaints such as make it better or less AI-looking. The generator cannot turn those into terms a model understands, so it will guess, and its guesses tend to be the generic quality words that do the least.

What are the most common mistakes with generated SD prompts?

• Pasting a 60-term negative block into every image. It dilutes the terms that matter, and on CLIP-based setups the important words risk falling past the 77-token cut.
• Writing sentences in the negative prompt. No people in the background puts people and background into the list. Write people, crowd.
• Using AUTOMATIC1111 weights on the hosted API. The documented ranges differ.
• Expecting negatives to work on Turbo at guidance 1. See the guidance section above.
• Changing three things between renders. You will not know which change helped.
• Trusting quality words. Masterpiece, best quality is a common habit in community prompts. Whether they help depends on the checkpoint, so test them rather than assuming.

Where Prompt Architects fits

The generator above is free and stays free. If you write image prompts often, Prompt Architects does the same formatting job inside ChatGPT, Claude and Gemini through a browser extension, with model-specific output for Stable Diffusion, Midjourney, DALL·E and Ideogram. Image prompt generation starts on the Pro plan. Saved library prompts can carry a separate negative prompt field, so a positive and its matched negative stay together; saving prompts needs Pro or above, since the Free plan does not include saved prompts. The Free plan does include 5 prompt enhancements per day, according to our FAQ page.

What we do not do is render images. You still need Stability's API or your own local setup for that, and nothing here replaces testing on your actual checkpoint.