A practical identity contract for product, character, and brand-led AI image workflows.
A reference image gives an image model a visual starting point. It does not tell the model which details are untouchable.
That distinction is easy to miss. An output can keep the broad idea of a product while changing its proportions. It can preserve a character's clothing and lose the face. It can borrow a package shape but invent a label. The image may still look polished, yet be unusable for the job that prompted it.
The fix starts before generation. Treat identity as an explicit contract: name the details that must survive, assign each reference a role, and review the result in passes instead of asking whether it simply "looks right."
Identity is a list of protected details
"Keep it consistent" is not a useful review standard. Consistent with what?
For a physical product, identity might include silhouette, dimensions, material, color, visible construction, packaging, and the relationship between parts. For a recurring character, it may include face shape, hair, age range, body proportions, clothing, and a small set of recognizable accessories. A brand-led image may depend on an approved palette, product family, type treatment, or composition rule.
Write those details down before choosing a style or scene. Separate them into three groups:
- Protected: details that must remain recognizable.
- Flexible: details the generation may reinterpret.
- Manually verified: logos, labels, prices, dates, legal copy, and other information that should never be trusted only because it appears in an image.
This turns a vague creative preference into something a reviewer can actually check.
Give every reference one job
Multiple references are useful only when each one adds a distinct instruction. Five unlabeled images can create five competing centers of gravity.
A clearer brief assigns one role to each source. One image defines the subject. Another provides composition. A third contributes material or palette. An environment reference should guide the setting without quietly replacing the subject.
You can write the assignment as a small piece of structured text:
reference_1:
role: subject identity
preserve: silhouette, proportions, material, primary color
reference_2:
role: composition
borrow: camera angle, negative space, subject scale
reference_3:
role: environment
borrow: lighting direction, background context
avoid: changing the subject design
This is a briefing pattern, not a special API format. Its value is that it forces the author to decide what each image controls.
The Pixonara multi-reference workflow follows the same principle: choose a compatible model based on the sources the job needs, then give those sources defined roles in the visual brief.
Separate preservation from transformation
A useful image-to-image prompt contains two instructions that are often mixed together: what should stay, and what should change.
Suppose the source is an approved product photo and the new output is a campaign image. The protected side of the brief might include the product's form, cap, material, and color. The transformation side might specify a low camera angle, dramatic side light, a wet studio surface, and clear space for final copy.
That can be written directly:
Use the supplied image as the source product.
Preserve the product silhouette, proportions, cap construction, material,
and primary color. Change the camera angle, lighting, background, and crop
to create a 4:5 campaign image. Leave the upper third visually quiet for
copy that will be added during final design.
Treat visible label text as a review item, not a guaranteed fact.
When one main source is being transformed, a focused image-to-image workflow is usually easier to reason about than a larger reference set. Add more sources only when the job truly has more visual roles.
Review the result in passes
A single overall judgment hides small failures. A better review moves from large structural details to smaller visible facts.
Start with silhouette and proportions. If the subject no longer reads as the approved product or character, there is little value in inspecting the lighting. Next check material, color, construction, and any distinctive parts. Then inspect labels, logos, generated text, hands, faces, and edges at full size.
Finish with composition and source interaction:
- Did the subject reference control the subject?
- Did the composition reference change framing without replacing identity?
- Did material or environment references leak unwanted objects into the result?
- Is the delivery crop usable in the intended channel?
- Which visible details still need replacement or manual verification?
Keep the original sources beside the output during this review. Memory is a poor comparison tool, especially after several attractive variations.
The Pixonara model directory documents current model-specific inputs and output controls. Those controls define what can be attempted; the identity checklist defines what can be approved.
Rejection is part of the workflow
Some outputs should not be repaired with a longer prompt.
If a protected shape has changed, a product variant has been invented, or several references have blended into the wrong subject, reject the image and simplify the brief. Remove unnecessary references. Reduce the number of protected details competing in one pass. Generate the core subject first, then build the environment in a later step when the workflow allows it.
Other failures belong in final design rather than regeneration. Exact campaign copy, legal language, prices, dates, and small packaging text are usually better added and verified outside the generator. The right question is not whether the model can produce something that resembles text. It is whether the final asset can be checked and approved with confidence.
A reference-led workflow succeeds when every important decision has an owner: the source image establishes direction, the brief defines the allowed change, and the reviewer decides whether the result is fit to publish.
