Z-Image, Seedream 5.0 Pro and Wan 2.7 Pro: which fits your task

"Image API" covers three different jobs: filling a folder with cheap drafts, delivering one finished frame from a long brief, and revising a picture you already have. Z-Image, Seedream 5.0 Pro and Wan 2.7 Pro split those jobs differently, and the differences sit in the request itself: tasks, output size, seed, reference count and default price. Which one fits depends on whether your input is a prompt, a picture, or both.

Short answer (based on our September and October 2026 tests). Pick Z-Image for new images from text when you want many cheap tries ($0.01 each). Pick Seedream 5.0 Pro for a finished 1k or 2k frame from a long, text-heavy brief, or for edits that combine up to ten references ($0.036 at 1k, $0.072 at 2k). Pick Wan 2.7 Pro when you need the biggest canvas (2k by default, 4k on text-to-image, a flat $0.075) or an edit steered by regions and a colour palette. For adult-oriented briefs, Wan 2.7 Pro's text-to-image task is the softened one: it toned down every level 3 and level 4 request. Its edit task, Z-Image and Seedream 5.0 Pro are not labelled that way, though Wan's edit task was still toned down at level 4.

Side by side

Z-Image: many cheap drafts of a new image

Z-Image has one task, text-to-image. At $0.01 a request, ten deliberate drafts cost $0.10, so you can compare one-clause changes at a low cost. Hold the seed to keep the starting condition fixed, but compare the actual outputs: it does not guarantee identical framing or isolate the effect of the edited clause.

The limits are real: no resolution field (each ratio resolves to a fixed size around one megapixel), prompts stop at 1,000 characters, and no reference image. Keep any text in the image to a few large words: it misspelled a mid-size italic line in every test run (see the label test below). Use it to find a composition, then take the wording to a larger model for the version you ship.

Leave enable_prompt_expansion off while you iterate: it rewrites your prompt and raises the price to $0.016. If you need exact pixel dimensions rather than ratios, that is a different Z-Image family: see the Z-Image Spicy workflow guide.

Seedream 5.0 Pro: a finished frame, or an edit from up to ten sources

Seedream 5.0 Pro takes 5,000-character prompts, offers fifteen ratios and two sizes, and is the only one of the three with a transparent-background option on edit. ByteDance's launch note positions it for infographics, posters and layout-heavy work, and says finer text rendering and pixel-level edit consistency still have room to improve, so check untouched regions of every edit.

Four details change cost or results:

• The default is 2k. If you omit resolution, you pay $0.072, not the $0.036 headline. Use 1k to confirm the composition, then render the delivery frame at 2k.
• There is no seed. Re-running gives a different image, and chained edits drift because each pass re-renders the whole frame. Edit again from the original references with one fuller instruction instead of stacking small ones. Wan behaves the same way: a seed narrows variation between runs but does not make an edit local.
• References cost extra. Every reference after the first adds $0.003 (ten references at 1k: about $0.063). Give each reference a stated job ("image 1 is the subject, image 2 only the room"); an unlabelled source gets blended in.
• Transparent output is narrow. background: transparent works only with PNG output and exactly one input image that is itself a PNG with an alpha channel. To pull one finished image apart into layers, use the separate layer-decomposition task in the layer-splitting guide. Cheaper family members are compared in the Seedream 5.0 side-by-side.

In our tests a two-image placement kept the clothing and room but lost the source person's likeness, so check the face against the source on every pass, on Seedream and Wan alike.

Wan 2.7 Pro: the large canvas, and edits steered by regions

Wan 2.7 Pro's text-to-image runs at 2k by default and goes to 4k, about four times the pixels, at the same flat $0.075. Choose it for a hero frame that will be cropped, enlarged or printed, or a layout where lettering needs the room.

The edit task accepts one to nine references, runs at 2k (4k is text-to-image only), has no default ratio (if you omit aspect_ratio the last reference sets the proportions), and adds bounding_boxes to mark where the change should happen. A box is a request, not a guarantee that everything outside it stays untouched, so compare the result with the source. color_palette takes three to ten colours (eight is recommended) and steers the palette of the whole frame, including set, wardrobe and light. For practical prompts and a worked edit, see the Wan 2.7 workflow guide.

Text inside the image: one label brief on all three

We sent one product-label brief (an amber honey jar with a large headline "WILDFLOWER HONEY", a small italic line "Raw & unfiltered" and a very small weight-and-batch line) to each model three times at 4:3. All three set the headline and the very small line correctly every time. The italic middle line separated them: Seedream 5.0 Pro spelled it correctly in 3 of 3 runs at 1k (and once at 2k), and Wan 2.7 Pro in 3 of 3 runs at its default 2k, while Z-Image misspelled it in all three. Wan's single 4k run also misspelled it, so a bigger canvas does not guarantee correct small text.

If the picture contains text you cannot fix by hand, use Seedream 5.0 Pro or Wan 2.7 Pro and check every size of lettering; if the text is a few large words, Z-Image's $0.01 frame is enough.

Content labels matter for adult-oriented work

The catalogue labels each endpoint by what it did on test prompts. Softened means the call succeeds and is billed but requests beyond the model's own limits can come back toned down. Check the label for the endpoint you will call, not the family name.

In these tests (review methodology), L1 is suggestive imagery, L2 moderate NSFW, L3 advanced NSFW and L4 higher-intensity adult scenes between adults. Each cell shows runs delivered as asked out of runs made; every other run was toned down (clothing added, key parts obscured or an unrelated image returned), and none was refused.

Wan 2.7 Pro text-to-image toned down all of its L3 and L4 runs, and its edit task, though labelled uncensored, was toned down at L4, so expect it to be less dependable than Seedream 5.0 Pro edit at the highest intensity. Detail is in the Wan tests, the Seedream 5.0 Pro test article and the Z-Image test article.

What a keeper costs

These are successful single-image prices. Failed tasks release the hold; tasks priced by output count settle for the images actually delivered. A successful image you cannot use still adds to the keeper cost.

The last column is a break-even. Z-Image is cheaper per keeper only when it gets you an acceptable image in fewer than about 3.6 tries if a Seedream 1k frame would be accepted first time, and about 7.5 against a one-shot Wan frame. The deliverables also differ (about one megapixel from Z-Image, up to 16 from Wan at 4k). Count billed outputs on a real brief, including successful images you discard. For monthly totals, use the API cost calculator.

Moving one brief between them

In the Playground first. Open the Z-Image, Seedream 5.0 Pro and Wan 2.7 Pro pages in turn. On each, open the Playground tab, paste the same prompt, pick 4:3 (one of the five ratios all three share), check the price estimate and run it. Compare the three results before you write any code.

Then the API, once you know which model you want. The input fields are not interchangeable, so here is the full Z-Image request:

curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"model": "alibaba/z-image-turbo/text-to-image",
       "input": {"prompt": "A ceramic teapot on a linen cloth, soft window light, muted palette", "aspect_ratio": "4:3", "seed": 42}}'

For Seedream 5.0 Pro change model to bytedance/seedream-5.0-pro/text-to-image, drop seed and add "resolution": "1k". For Wan 2.7 Pro use alibaba/wan-2.7-pro/text-to-image and add "resolution": "2k". Fields, results and errors are in the API reference; the starter guide walks through a first request.

Let an Agent do it. An agent can make the same call through the CLI, the MCP server or the Agent Skill. All three read the live model schema, and CLI and MCP show the price and ask you to confirm before creating a task. With SPICY_API_KEY set:

# CLI
npx @spicyapi/cli tasks create --model alibaba/z-image-turbo/text-to-image \
  --input-json '{"prompt": "A ceramic teapot on a linen cloth, soft window light, muted palette", "aspect_ratio": "4:3", "seed": 42}' --wait

# MCP server: the command your assistant's MCP config runs
npx --yes --package=@spicyapi/mcp spicyapi-mcp

# Agent Skill
npx skills add spicyapi-ai/spicy-skill

Questions people ask

Is Z-Image Turbo better than Seedream 5.0 Pro? Not as a blanket statement. Z-Image is far cheaper and suits a draft loop of new images; Seedream 5.0 Pro takes up to ten references, longer prompts and a 2k delivery size. For specs side by side, see the Seedream 5.0 Pro vs Z-Image page.

Does 4k from Wan 2.7 Pro mean a better image? Not by itself: it is about four times the pixels of 2k at the same price, but Wan's 4k label frame misspelled text its 2k frames spelled correctly. Edits stay at 2k.

Can I swap only the model name in an existing request? Only after removing the fields the new endpoint does not list and re-checking the allowed values. Each compare page covers one task: Wan 2.7 Pro vs Z-Image shows text-to-image, while Seedream 5.0 Pro vs Wan 2.7 Pro shows image editing only.

How to choose

Exploring new images from text: Z-Image. A long or text-heavy brief, or an edit that combines several references: Seedream 5.0 Pro, drafted at 1k first. A 4k frame, or an edit steered by regions and a colour palette: Wan 2.7 Pro. Then run one real brief through your shortlist and count the tries you needed: the accepted image, not the request price, is what you pay for.