Design exploration
A designer can keep the framing from a reference while replacing the subject and palette. Read the fields as a creative checklist, then decide which relationships belong in the next concept.
Image to JSON prompt.
Separate a visual reference into clear, editable fields.
A visual brief you can take apart. Separate the scene into editable fields, then change only what matters.
YOUR CREATIVE WORKSPACEREFERENCE → PROMPT
01 / Your referenceNO SIGN-UP
Drop an image. Find its words.or click to choose a fileJPG, PNG, WebP · Up to 10 MB · Paste with Ctrl / ⌘ + V
The image host must allow browser access. If blocked, download and upload the file instead.
Up to 5 connected attempts per tool each UTC day. Remaining usage appears after a request.
Processed in memory. Uploads are never saved by this site. Privacy details
02 / Your promptAWAITING REFERENCE
→Aa
A picture is a starting point.
Your prompt will appear here, ready to edit, copy, and make your own.
A creative interpretation, never an exact reverse of an original prompt.
THREE CLEAR STEPSFrom a starting point
From a starting point
to your next creation.
- 01
Add a reference
Upload, drop or paste a JPG, PNG or WebP you are allowed to use. A direct image URL also works when the host permits browser access.
- 02
Keep JSON selected
Choose a writing target, then generate. Subject, scene, composition and lighting appear as separate fields rather than a single paragraph.
- 03
Review and reuse
Check each value against the reference. Edit the result, copy the JSON, and adapt the schema to the application where you plan to use it.
AN ILLUSTRATIVE EXAMPLESee the brief.
See the brief.
Make the decisions.
This original writing example is not a live analysis of an uploaded reference.
INPUT
EXAMPLE
EXAMPLE
An illustrative reference: a matte green cup on pale stone, with soft light from the left. Original image pending: TODO：用生图工作台生成。
THE EDITABLE OUTPUT
{"subject":"A matte green ceramic cup","scene":"Pale stone surface","composition":"Off-center, open space to the right","lighting":"Soft light from the left","text":"No added text","aspect_ratio":"3:2"}
A brief you can work with.
01 / A USEFUL DETAIL
Find the field you need
Separate composition from subject and lighting from style. A readable structure helps you make a small change without rewriting every part of the brief.
02 / A USEFUL DETAIL
Use your writing target
GPT Image 2 stays the default, with GPT Image 2.5 and other image-writing presets available. A target changes the brief, not the vision provider.
03 / A USEFUL DETAIL
Keep control of the result
The output remains editable. JSON is a convenient container for your description, not a claim that every image application accepts the same request schema.
WHEN IT HELPS
A place in your workflow.
Repeatable briefs
A maker maintaining several product concepts can compare versions field by field. Save your own copies in an appropriate workspace; Aizvu does not create a project library or retain a history of your uploads.
A PRACTICAL FIELD GUIDEFewer assumptions.
Fewer assumptions.
A clearer direction.
A schema for inspection
The JSON format uses subject, scene, composition, lighting, style, text, colors and aspect ratio to organize a reference. These are Aizvu editorial fields, chosen because they are easy to inspect. They are not a hidden image-model language, a required vendor format or a complete API request. When you copy the object into your own application, check which fields that application understands. You may need to combine values into natural language or map them to your own schema. Keep the meaning of each field clear instead of assuming that a shared field name guarantees compatibility.
Separate observation from a decision
A photograph can show where a subject sits and how a surface appears. It cannot establish an exact lens model, manufacturing material or original generation history. Read every field with that distinction in mind. If a detail is unclear, keep the wording uncertain rather than inventing precision. Once you deliberately change the subject or scene, the new value is a creative decision. That is useful, but it should not be confused with a fact extracted from the original image. The resulting object is a brief you can reason about, not a forensic record.
Change one relationship at a time
Begin with the composition field if the arrangement is what attracted you to the reference. Preserve the relative subject size, open space and viewing angle while changing the item itself. Then review lighting separately. A different subject may need a simpler description, and a busy palette may work better when reduced. Keep a copy of the version you liked before you make another change. This is an editing practice rather than a promise of model consistency: your chosen generation tool can still interpret the same brief differently on each attempt.
Treat visible lettering carefully
Inspect quoted words against the reference, especially when the image contains small text, stylized characters or a cropped label. Do not accept a plausible-looking phrase simply because it fits the scene. Unclear lettering should remain unreadable or Unconfirmed. For a new design, write your own exact wording and explain its placement in the text field. Describe alignment, contrast and approximate scale without pretending to identify an exact typeface. After generating an image elsewhere, inspect the rendered letters again; a carefully written JSON value does not guarantee perfect spelling.
Keep your JSON valid
The connected backend checks that JSON output parses as an object before returning it. Once you edit that output in the workspace, you can introduce a missing comma, an unescaped quote or an accidental extra bracket. Check syntax again before using a manually changed object in another application. If you only need a brief to paste into an image tool, detailed natural language may be more convenient. Choose the format for your workflow rather than treating structured output as inherently more accurate. The same uncertain visual interpretation can appear in either format.
Understand the request boundary
Aizvu compresses the image in your browser, then sends it through its backend in request memory. The application does not intentionally save uploaded images or generated objects to a file library. Connected requests reach the configured vision provider, whose retention policy also applies. Use a reference that is appropriate to share and avoid sensitive material. In local example mode, the object describes an illustrative cup rather than your actual upload. That label matters: it lets you test copying, editing and field layout without mistaking a sample for a genuine image analysis.
GOOD TO KNOW
Before you create.
01Is this an official JSON schema?+
No. It is an Aizvu creative-brief schema. Adapt it to the receiving application, and keep API settings separate from descriptive fields.
02Does JSON make the interpretation more accurate?+
No format guarantees accuracy. JSON makes individual decisions easier to inspect, while the underlying image interpretation can still be uncertain or incomplete.
03Can I edit the object?+
Yes. The result is editable before copying. Check syntax after manual changes, and do not assume every image tool accepts raw JSON.
04Which files and targets are supported?+
Use JPG, PNG or WebP up to 10 MB. GPT Image 2 is the default writing target; GPT Image 2.5 and the other listed image presets are optional.
05Are images stored or sent elsewhere?+
Aizvu processes uploads in memory without intentionally saving image files. Connected requests reach the configured vision provider, whose policies also apply.
06Why do I see an example?+
Without a configured vision service, local development shows a labelled illustrative result. It does not analyze the uploaded reference. Connected production also requires security configuration.