Choose an AI product-image model by the operation you need: preserve a packshot, replace a background, compose several references, create a transparent asset or invent a new scene. These are different tasks. A strong text-to-image result does not establish that the same model will preserve your packaging in an edit.
The most useful shortlist is a table of required controls followed by a small acceptance test. This guide uses models and operations checked on 19 October 2026, with Raywake links to the relevant workflows.
Match the task to the operation
| Your task | Controls to look for | Candidates to inspect |
|---|---|---|
| Change a packshot's environment | Image-edit input and clear preservation instruction | Nano Banana 2.1 Edit, Seedream 5 Edit |
| Change a bounded area | Reference image plus mask | GPT Image 2.5 Edit |
| Combine product, set and material references | Multi-reference editing | FLUX.2 Pro/Max, GPT Image 2.5 Edit |
| Produce a cutout with alpha | Explicit transparent-background control and compatible output format | GPT Image 2.5 |
| Build a text-heavy product graphic | Text-layout capability plus exact-copy review | Qwen Image 3 Edit |
| Draft a fictional product scene | Text-to-image operation | A current generation endpoint matching the required size |
These are capability-based candidates, not a quality leaderboard. Start with the requirement that would disqualify an output, then choose a model that exposes the appropriate control.
What must remain unchanged?
Create a preservation list before generating. For a cosmetics bottle it may include silhouette, cap shape, logo, label copy, fill level and color. For footwear it may include panel seams, sole shape, fasteners and material texture.
Separate those features from the parts you want to change. A background replacement should not redesign the package. A new camera angle may require the model to infer surfaces that the reference does not show; decide whether that inference is acceptable for the asset's purpose.
When accuracy is essential, use the approved product photograph as the final product layer and generate the surrounding scene separately. That workflow can preserve detail that an unrestricted full-image regeneration might alter.
References, masks and transparency solve different problems
A reference tells a model what something looks like. A mask identifies a region for an edit. An alpha channel lets the resulting asset sit over another background. None of those features automatically implies the other two.
For example, GPT Image 2.5 Edit exposes references and an optional mask. Its background control can request transparency. Nano Banana 2.1 Edit accepts image references, but the reviewed contract does not expose the same explicit transparent-background switch.
Sources: GPT Image editing contract, Nano Banana editing contract. A prompt that says “transparent” still needs an alpha-channel check before delivery.
Keep the reference set small and purposeful
Use one authoritative image for product identity. Add a second reference for the environment or lighting only if it answers a specific question. State which image supplies which element.
For more complex compositions, FLUX.2 reference workflows can combine several sources, while the reviewed GPT Image edit contract accepts up to sixteen references. The maximum is a limit, not a recommendation to fill every available slot.
If the model blends two products or borrows a label from the wrong reference, simplify the set. Changing three inputs and the model at once makes the next result difficult to interpret.
Price a defined output, then review it
Choose output dimensions, format, quality and count before obtaining a Raywake quote. A 1K concept, a large edit with multiple references and a transparent cutout are not equivalent pricing units.
Use a small evaluation record:
- Required task and details that must remain unchanged.
- Model, exact settings and source files.
- Quoted amount and eventual settled charge.
- Pass or fail for product identity, copy and framing.
- Reason for any deliberate retry.
For a batch, divide the total generation spend by approved assets. This keeps rejected variations visible. It also prevents a low headline price from hiding a workflow that needs many corrections.
Plan for the next production step
If the image will become a video, approve the still first. Keep space for the intended camera move, avoid cropped limbs or product edges, and choose the eventual video aspect ratio early. Read the first/last-frame and reference-video guide before deciding how that approved still should guide motion.
For a static ad, leave copy editable when possible. For a product listing, check the required background, dimensions and disclosure rules of the destination before making the final export.
Common questions
Which model is best for every product?
This guide does not make that claim. The right candidate depends on the edit, input controls and acceptance criteria.
Can a higher resolution repair an inaccurate label?
No. It may produce a larger inaccurate label. Correct the source, instruction or compositing workflow first.
Does Raywake require separate balances for these models?
The studio and API use the same account USD balance. Check current pricing and the quote for the selected operation.
Choose a product-image workflowReviewed on 19 October 2026. The cover is an editorial illustration generated with Nano Banana 2; it is not a test result from the models discussed.