Which Image Model Should You Use? A Practical Seedream and Wan Guide

The wrong question is “which image model is best?” The useful question is: what must remain true in the finished image? A product shape, a brand colour, Chinese text, a recurring character, a rough creative direction, or a low-cost first draft all lead to different choices.

Our 2026-07-20 gateway-reconciled snapshot contains four current image SKUs across three families from the wider routing study: doubao-seedream-5-0-260128, doubao-seedream-4-5-251128, wan2.7-image, and wan2.7-image-pro. The wider study considered additional families; this public edition deliberately limits itself to models present in the catalog snapshot.

That scope matters. This is a selection guide for models you can call here today, not a claim that one vendor has won every image task.

The short answer

• Pick Wan 2.7 Image when an existing product, character, or brand asset must survive the generation. Reference-driven work is a control problem before it is an aesthetics problem.
• Pick Seedream 5 for a new campaign visual or a social series where the creative direction is still being explored. It is the better first route when you are designing the composition rather than preserving an asset.
• Use a standard model for exploration, then pay for the higher-control route only after a direction has been selected. Re-running a cheap image many times is not necessarily cheap; measure the cost of a usable image.

Do not use a generic “flagship” default for every request. It hides the real trade-off: an image that looks impressive but changes a SKU, a logo, or a character is a failed production asset.

What is actually available in the current catalog

The prices above are the current displayed service rates from the gateway-reconciled catalog, not a forecast of vendor list prices. Size, reference inputs, retries, and the chosen upstream configuration can change the actual cost of a finished asset. Check live pricing before a production run.

A routing rule that survives real work

Start with the asset you cannot afford to lose.

This is intentionally a task rule, not a vendor ranking. A beautiful prompt-only image does not prove that the same model will preserve a real product in ten variants. Conversely, a reference-oriented route is unnecessary overhead for a disposable concept board.

The metric that changes the decision: usable-image cost

Model price is only one part of production cost. The more useful metric is:

usable-image cost =
  (all generation attempts + editing/review time) / accepted images

That formula is why a nominally cheaper option can lose. If a product image drifts and needs three retries plus manual cleanup, a single higher-control generation may be cheaper in the only unit that matters: an asset your team can ship.

Here is a repeatable evaluation card for your own work. Run the same task through the candidate models, keep the input assets and output specification fixed, and record the result rather than relying on a memorable one-off image.

{
  "task_id": "sku-jar-01",
  "task_type": "product_reference_edit",
  "reference_assets": 2,
  "output_spec": "1:1 marketing image",
  "model": "wan2.7-image",
  "attempts": 3,
  "accepted": 1,
  "product_shape_preserved": true,
  "text_or_logo_review": "manual",
  "generation_cost_usd": "record actual billed amount",
  "repair_minutes": 8
}

The image-routing brief defines a 50-task candidate set: Chinese posters, ecommerce products, brand or IP characters, cross-border ads, and editing tasks. It is a test plan, not a published benchmark. Until repeated runs exist, do not turn the table above into numerical quality scores.

Rights and reference images are part of routing

An image API accepting a reference file does not grant permission to use the person, product, logo, or character in that file. Treat rights metadata as an input to selection:

asset type → rights owner → consent status → allowed use → approved model path

For a fictional illustration, normal safety review may be enough. For an employee, customer, contracted talent, product photograph, or trademarked asset, retain the right to use the source and the intended commercial use before generation. Do not use a model comparison as a way to bypass a provider's safety checks or a subject's consent.

What we have not measured yet

We have not published repeated ChinaAPI benchmark results for text accuracy, product preservation, character consistency, latency, or acceptance rate across these models. That is a real limitation, not a footnote. The routing recommendations above are based on model capabilities, the gateway catalog, and the production failure modes the study identifies—not a claim of measured superiority.

Use this guide to choose the first test. Then keep the task card, compare usable-image cost, and promote a model to your default only when your own accepted-output data supports it.