Everything is weighted
Nested parentheses and competing values make priority hard to see. Return to plain language, then add one emphasis where the baseline consistently underplays an important concept.
Stable Diffusion prompt lab
Stable Diffusion Prompt Generator — Positive & Negative Prompts
Translate a reference into a Stable Diffusion prompt with a clear positive description, an editable negative direction, and visible cues for composition, lighting, material, camera feel, and style. Use the result as a checkpoint-aware starting point, not a claim about hidden source settings.
- Structure
- Positive + negative
- Control
- Weights and settings
- Request
- 1 credit
Reference translated
Conditioning / exclusions(minimal product photography:1.2), white wristwatch, overhead centered composition, brushed silver case, matte silicone strap, cool gray surface, diffused studio light, crisp material detail
Positive
subject, scene, light, camera
Negative
text, logo, clutter, blur
Weight
(centered composition:1.15)
Outside text
seed, steps, CFG, sampler
Stable Diffusion input
Extract the visual brief before choosing model settings
Generate the six outputs, open the Stable Diffusion prompt, and keep the visible scene description separate from checkpoint, sampler, CFG, seed, and structural controls.
Drop images herePaste an image, or choose how you want to add one.JPG, PNG, or WebP · up to 10 MB each
No image selected
Add an image to create all six prompt formats.General · Midjourney · FLUX · Stable Diffusion · ChatGPT Image · Nano Banana
Portable text, local context
A Stable Diffusion prompt is one part of a larger generation setup
Text conditioning describes the image you want. The surrounding Stable Diffusion workflow can also include a checkpoint, VAE, sampler, scheduler, steps, CFG value, seed, LoRA, ControlNet, reference adapter, canvas size, and other node or interface settings. A well-written prompt matters, but it does not control the full system by itself.
This generator reads visible content and converts it into a positive description plus useful negative guidance. The positive text explains the subject, environment, composition, light, color, camera feel, materials, style, and mood. Negative guidance identifies common output directions that would conflict with the reference, but it should remain short enough to understand and revise.
The generated text cannot recover an original text instruction. A flattened image does not expose the checkpoint, seed, CFG, sampler, LoRA weights, inpainting history, ControlNet map, upscaler, or post-processing steps that produced it. Even metadata can be stripped or rewritten. This page therefore reconstructs a practical brief from visual evidence rather than presenting speculation as provenance.
One analysis also produces General and four other platform-specific versions. This version is separate because token emphasis, negative conditioning, checkpoint vocabulary, and external generation settings create a different editing workflow. Copying a conversational ChatGPT instruction or a parameter-heavy Midjourney phrase into a Stable Diffusion interface can work, but it hides which control is responsible for the result.
Positive and negative fields
Separate the Stable Diffusion prompt from the settings around it
The prompt should describe the target image. Generation settings should remain visible as settings. Keeping that boundary clear makes experiments reproducible and prevents an unreadable string from becoming the only record of your workflow.
- 01
Positive subject
Begin with the main subject and the details that define its silhouette, age group, clothing, surface, or orientation. Put important concepts before decorative polish. Many interfaces and models react differently to token order, but early clarity remains easier for a human to audit.
- 02
Scene and composition
Add environment, camera position, framing, and object relationships. The description should distinguish a centered overhead product image from an eye-level lifestyle photograph. Name foreground and background anchors when they create depth, and use reference-control tools when exact geometry exceeds what text can hold.
- 03
Light and material
Describe the source, hardness, direction, shadow behavior, and reflections that make the material readable. “Large diffused softbox, controlled silver reflections, soft contact shadow” gives the model more usable evidence than “premium studio lighting.”
- 04
Selective emphasis
Weights can strengthen or weaken a concept in interfaces that support the syntax. Use them sparingly and document the format your front end expects. One weighted composition phrase can be diagnosed. Weighting every token makes the instruction hard to read and can create brittle results.
- 05
Negative direction
Keep the negative conditioning tied to visible risks: unwanted text, duplicate objects, clutter, blur, distorted anatomy, or a conflicting medium. Do not paste an enormous inherited list without reading it. Some terms can remove details you actually need or behave differently across checkpoints.
Checkpoint interpretation
Checkpoint choice changes what the same words mean
Stable Diffusion is used through many checkpoints and interfaces. Text that works with a photoreal checkpoint may look flat in an illustration checkpoint, while a style tag learned by one model may have little effect in another. Record the checkpoint and major adapters beside the instruction. Otherwise a useful sentence can appear broken when the model context has changed.
Negative prompts also vary by workflow. Stability AI APIs expose a negative_prompt field for supported generation endpoints, and older SDK documentation describes negative weighting as a way to discourage concepts. Local interfaces may parse emphasis syntax or embeddings differently. Check the software you are actually running instead of assuming one universal prompt grammar.
CFG, steps, sampler, seed, and canvas dimensions belong outside the positive text. CFG changes how strongly generation follows the conditioning. A seed can help recall a result only while other relevant settings remain stable. These controls can explain a large visual change, so save them beside the prompt when reproducibility matters.
Text also has limits. The prompt can suggest a top-down layout, but ControlNet, an image adapter, or an image-to-image workflow can carry structure more directly. Use text for semantic and aesthetic direction. Use the appropriate conditioning tool for pose, depth, edges, identity, or exact reference influence.
Known-model testing
Debug a Stable Diffusion prompt one conditioning layer at a time
The easiest way to lose an afternoon is to change prompt text, checkpoint, LoRA weight, CFG, sampler, and seed together. A controlled pass keeps the cause of each change visible.
- PASS 01
Choose a known model context
Select the checkpoint and required adapters before judging the generated text. Use a canvas that matches the reference orientation. If you are comparing two revisions, keep those model settings stable so the difference has a meaningful cause.
- PASS 02
Run the positive prompt first
Inspect subject, composition, light, and medium with a restrained negative field. If the scene fails, correct the positive description before expanding exclusions. A negative list cannot supply a missing camera position or tell two objects where to sit.
- PASS 03
Add targeted negatives
Use the first output as evidence. If unwanted typography appears, discourage text and logos. If duplicate objects recur, name that failure. Keep the negative list specific enough that you can explain why each term is present.
- PASS 04
Move structure to structure controls
When wording changes do not hold pose or geometry, stop stretching the text. Use image-to-image, ControlNet, depth, edges, pose, or a suitable reference adapter. Language is one conditioning channel, and the workflow improves when each channel carries the information it handles well.
Portability failures
Why a Stable Diffusion prompt stops being portable
The same text can change under a different checkpoint, parser, embedding set, or workflow. Portability starts with readable language and an honest record of the settings that live outside it.
The negative field is inherited
Large copied lists may contain terms for a different model, style, or anatomy problem. They can suppress texture, color, or detail needed by the reference. Keep only the negative terms you can justify.
Settings masquerade as prompt quality
A new checkpoint or LoRA can improve an output while the text remains unchanged. Save the context. Do not attribute every change to the Stable Diffusion prompt when other conditioning changed at the same time.
Text is asked to lock geometry
Longer text does not guarantee an exact pose, face, edge map, or object count. Use a structural control when the workflow requires it. The prompt should remain the semantic brief rather than an overloaded substitute for every node.
Checkpoint and syntax questions
Stable Diffusion prompt questions
What is a Stable Diffusion prompt?
A Stable Diffusion prompt is text conditioning that describes the image you want. It usually works alongside a model checkpoint and generation settings. Some workflows also use a negative prompt, weights, LoRAs, ControlNet, reference adapters, or image-to-image conditioning.
Can this recover the original Stable Diffusion prompt?
No. Pixels do not reveal the original Stable Diffusion prompt or the complete workflow. The checkpoint, seed, sampler, CFG, steps, adapters, node graph, edits, and metadata may be unknown. The result is a new visual reconstruction.
Do I need a negative Stable Diffusion prompt?
Not every generation needs a long negative field. Start with the desired image and add targeted exclusions when a repeated failure appears. Check your model and interface because negative conditioning behavior differs across workflows.
Should I put CFG and steps inside the Stable Diffusion prompt?
No. Keep CFG, steps, sampler, seed, size, and checkpoint in the generation settings or workflow record. The Stable Diffusion prompt should remain readable as the semantic description of the image.
Will the same Stable Diffusion prompt work in SDXL and SD3?
The broad visual idea can transfer, but token response, preferred detail level, negative conditioning, and available controls can differ. Test the prompt in the exact model, keep settings visible, and revise from the output rather than assuming identical behavior.
Is this Stable Diffusion prompt generator free?
Visitors receive three daily credits. One image uses one credit and generates all six outputs, including the model-specific version. Signed-in users can edit and save History. Pro subscribers can submit multiple images in a batch.
Test the Stable Diffusion prompt in a known model context
Start with a readable Stable Diffusion prompt and a known workflow
Upload the reference, correct the positive Stable Diffusion prompt, and add only the negative terms supported by output evidence. Keep checkpoint, seed, sampler, CFG, steps, adapters, and structural controls beside the text. A Stable Diffusion prompt is only half of a reproducible record; save the runtime settings with it. Compare each revised Stable Diffusion prompt against the same baseline before changing another control.