The greatest value of an image‑to‑image reference isn’t to have the AI “copy exactly,” but rather to transform abstract concepts—such as subjects, poses, compositions, or visual directions that are hard to articulate in words—into visible constraints. To ensure stable results, your prompt must clearly address two key questions: which elements must remain unchanged, and which can be modified.

The accompanying images are original illustrative materials created for this tutorial, generated by Codex; they do not represent a comparative review of PixPix models or actual platform testing. The PixPix feature descriptions are based on currently available public pages; input limitations, dimensions, and options may vary across different models and current interfaces—please refer to the actual interface for accurate information.

Quick conclusion: First list what must stay the same, then specify what can change.

An executable image‑to‑image request can be condensed into this formula:

Reference image character + Required identity/structure/composition to retain + Objects allowed to change + Desired scene or style + Lighting and proportions + Prohibited additions

If a single prompt simultaneously asks for changes in characters, actions, camera angles, backgrounds, and styles, the model will struggle to determine which aspects should serve as anchors. A more reliable approach is to modify only one major variable at a time, while continuing to use the previously confirmed image as a reference.

What exactly can a reference image control?

Currently, PixPix’s publicly available image‑to‑image entry point supports uploading reference images, sketches, or screenshots, enabling the model to generate new images based on the original image’s structure, composition, and style—ideal for style expansion, character transformations, product scene extensions, and visual exploration.

Subject and identity

Facial features, hairstyles, clothing, body types, or product outlines, colors, materials, and key structural details all fall under the category of unchanging subject attributes. Avoid simply stating “keep consistent”; instead, explicitly list each observable characteristic.

Pose and composition

Reference images make it easier than text to convey a person’s positioning, gaze direction, camera height, proportion of the subject, and relationships between objects. If these relationships must remain intact, clearly state: “Maintain the same pose, camera height, subject scale, and framing.”

Style and lighting

A reference image can also serve solely as a visual guide—for color, texture, material, lighting, or layout. In such cases, specify that it is a “style reference” to prevent the model from inadvertently including people or products from the reference image in the final result.

Original case study: Urban cyclist Lin

In this example, we create a fictional character named Lin, with the following defining traits:

• East Asian woman, around her twenties, with medium-length black hair;

• Wearing a lightweight cobalt-blue windbreaker, dark gray straight-leg pants, and white low-top shoes;

• Matte coral-colored cycling helmet;

• Holding onto a silver city bicycle;

• Both the entire figure and the bicycle are fully within the frame, with a fixed viewpoint and line of sight.

This master image encompasses identity, attire, pose, props, and composition. If you later wish to change only the style, avoid altering the background layout; if you want to switch scenes, refrain from modifying clothing or movements at the same time.

Procedure: Break down the request into one variable at a time

Step 1: Assign a role to each reference image

When uploading multiple images together, don’t leave it up to the model to guess:

• Reference Figure A: Subject’s identity and attire;

• Reference Figure B: Pose and composition;

• Reference Figure C: Color, material, or illustration style.

Even if there is only one image, you should still specify in your prompt what it represents. The clearer the character, the fewer conflicts will arise.

Step Two: Write down verifiable invariants

In Lin’s case, the invariant is not “the same person,” but rather:

Keep the subject’s facial features, black medium-length hair, cobalt-blue coat, coral-colored helmet, dark gray pants, white shoes, posture with hands resting on the handlebars, silver bicycle structure, overall composition, and viewpoint unchanged.

For products, this can be replaced with: “The bottle’s length-to-width ratio, cap structure, label placement, material, color, and branding remain unchanged.”

Step Three: Describe only the changes for this round

First decide which category this round belongs to:

• Change only the rendering style;

• Change only the background and time;

• Change only clothing or a single object;

• Expand only the canvas;

• Correct only localized errors.

Do not try to complete all these tasks at once. First obtain a version with stable structure, then proceed to the next round.

Example One: Change only the style

Example instruction:

Convert the reference image into a contemporary woodblock-printed editorial illustration. Keep Lin’s facial features, body proportions, expression, cobalt-blue coat, coral-colored helmet, dark gray pants, white shoes, standing pose, silver bicycle, camera angle, and cropping unchanged; only alter the rendering style. Use cobalt blue, coral, dark gray, and warm paper tones, preserving halftone patterns and subtle overprinting textures. Do not add any figures, text, or logos.

Here, the key verb is “only change the rendering style.” While style goals can be specific, do not simultaneously request a new scene or redesign of clothing.

Example Two: Change only the background and lighting

Example instruction:

Transform the environment into a riverside bike path after a light rain, bathed in blues of twilight. Only modify the background and ambient lighting; keep the subject’s identity, hairstyle, coral-colored helmet, cobalt-blue coat, dark gray pants, white shoes, body proportions, hand position, bicycle structure, pose, overall scale, and framing unchanged. The road surface has a slight wet sheen, distant city lights appear softly blurred, and no pedestrians or other bicycles are added.

When “background replacement” does not even allow changes to the subject’s lighting, the result often ends up looking like a simple sticker. A more reasonable approach is to permit ambient lighting to influence the subject while still locking the subject’s identity, clothing colors, and structural details.

When to overhaul the entire image, and when to perform localized edits

If the overall direction is off—for example, if the style, shot, or setting completely fails to meet expectations—re-generating the whole image is usually more efficient. However, if 80% is already correct and only the hands, edges, text, or a small object need adjustment, use PixPix’s publicly available local retouching or image-editing capabilities instead, avoiding a full rework that could introduce new deviations.

PixPix page for GPT Image 2The public description states that you can start from a prompt or reference image, and it’s suitable for product images, posters, typography design, and cross‑size visual series;Meanwhile, the PixPix page for Seedream 5.0 Proemphasizes clearly defining objectives, actions, layouts, styles, lighting, and invariants within editing instructions. The specific choice should be determined based on the current task and actual interface capabilities, rather than assigning a fixed model to every project.

Common mistakes and how to correct them

Mistake #1: Simply writing “the same as the original”

“Same” doesn’t specify priorities. List facial features, clothing, product structure, poses, or camera angles one by one so the model knows which changes constitute failure.

Mistake #2: Conflicting reference images

One image provides a person, another gives a pose, while a third introduces different clothing and characters—leading easily to confusion. Assign clear roles to each image and explicitly ignore any elements irrelevant to the task.

Mistake #3: Changing too many variables at once

First perform style transfer, then switch scenes, and finally make localized adjustments. Keep only one confirmed result per round, so if something goes wrong, you’ll know exactly which step caused the deviation.

Mistake #4: Leaving small text and logos entirely to generation

Product labels, promotional prices, and brand logos must be checked word by word. Once the overall structure is correct, prioritize making localized edits to text areas instead of recreating the entire image just for a single character.

Mistake #5: Relying solely on thumbnails

Zoom in to examine faces, fingers, wheel spokes, product edges, reflections, contact shadows, and text. Details that appear natural in thumbnails may become distorted when viewed at full size.

FAQ

Is more reference imagery always better?

Not necessarily. Each image should have a clearly defined role. If two images conflict regarding characters, composition, or style, increasing their number will actually reduce controllability.

Will image‑to‑image generation simply copy the original?

Image‑to‑image generation reads structural, compositional, postural, or stylistic information and regenerates the scene accordingly—it should not be seen as merely copying pixels. Before using images that don’t belong to you, always verify the scope of authorization and avoid requesting reproductions of protected, unique works.

Why do characters still change faces?

Reference image clarity, angle variations, model capabilities, and the number of variables modified at once all affect identity stability. Use high‑resolution master images, clearly define facial and clothing features, minimize simultaneous variable changes, and iterate starting from confirmed results.

When is text‑to‑image generation more appropriate?

When you only need to explore entirely new directions without requiring specific characters, products, poses, or layouts, text‑to‑image offers greater freedom. However, whenever your task specifies “maintain this person, this product, or this composition,” it’s better to start with reference images.

Summary: Treat reference images as constraints, not mere decoration

The stability of image‑to‑image generation stems from clear task boundaries: what the reference image is responsible for, which elements must remain unchanged, what will be modified in this round, and how final acceptance will be determined. With a master template and a list of invariants prepared, you can proceed to PixPix’simage‑to‑image and image‑editing entry, beginning with a single‑variable adjustment.