guide
How to Use Seedream 5.0 Pro for Multi-Reference Image Generation on EvoLink
Jessie
COO
July 7, 2026
Updated on July 8, 2026
12 min read
Updated July 8, 2026. Facts in this guide are verified against the live EvoLink route. The maintained price table and playground live on the Seedream 5.0 Pro page; this guide covers the workflow, not the price list.
Quick Answer: Use Seedream 5.0 Pro When References Matter More Than Volume
If your image task depends on reference images — a product that must look like your product, a character that must stay the same character across ten assets, a brand style that must survive regeneration — Seedream 5.0 Pro is the route built for that job. It accepts up to 10 reference images per request, generates one high-fidelity image at a 1K or 2K tier or exact pixel dimensions, and supports instruction-based editing on source images.
If instead you need high-volume drafts, story sets (multiple related images in one request), or the lowest cost per image, start with Seedream 5.0 Lite — it is cheaper and supports sequential sets, which Pro does not.
You can try everything in this guide interactively in the playground, then move to the API with the same parameters. To choose between Seedream family routes first, see the Seedream model family.
What This Guide Covers and What It Does Not Replace
Covered here: how to organize reference images so the model uses them the way you intend, how to pick size and output format safely, prompt patterns that assign a role to each reference, cost mechanics you should plan around, and the guardrails to set before this goes into a production UI or backend.
Not replaced by this guide: the Seedream 5.0 Pro page remains the canonical source for current prices and the live playground; the API reference on that page documents the full request contract; and the family comparison is the right place to decide which Seedream model to use. This guide assumes you have decided references matter and want to use Pro well.
Confirmed Seedream 5.0 Pro Facts for EvoLink Users
Verified as of July 7, 2026 against the EvoLink route:
| Capability | Current state | What it means for your product |
|---|---|---|
| Reference images | Up to 10 per request via image_urls | Enough for subject + style + background + brand assets in one call |
| Input image billing | Each input image 0.1734 credits | A single-reference edit flow adds one input-image charge |
| Output | One image per request (n = 1) | Fan out N requests for N variations; do not build UI around multi-image responses |
| Output tiers | 1K, 2K, or exact WIDTHxHEIGHT (1024x1024 up to 2048x2048) | Generate at placement size; skip the crop step |
| Tier boundary | Outputs above 2,360,000 total pixels bill as 2K | An exact-size request can silently cross into the 2K tier — check the math |
| Output format | jpeg (default) or png | png for transparency and post-production pipelines |
| Not supported | Sequential image sets, streaming, web search tools, guidance_scale, layers, prompt_priority | Hide these controls in your UI; requests that need them belong on other routes |
| Task model | Asynchronous task with polling or callback | Design for eventual results, not synchronous responses |
Two of these deserve emphasis because they are the most common integration surprises:
n is fixed to 1, and sequential image sets are not supported on Pro — both are supported expectations users carry over from other Seedream routes.Step 1: Choose the Right Seedream Route Before You Generate
Multi-reference work is Pro's job, but not every image task is multi-reference work. Route first, generate second:
| Your task | Use Pro | Use Seedream 5.0 Lite | Use another route |
|---|---|---|---|
| Product/character must match references exactly | ✅ | ||
| Instruction-based edit of an existing image | ✅ | ||
| Exact placement dimensions (e.g. 1376x768 banner) | ✅ | ||
| High-volume drafts, cost-first | ✅ ($0.031/image) | ||
| Story sets / sequential related images | ✅ (sequential supported) | ||
| More than 10 references in one request (up to 14) | ✅ | ||
| 4K output required (above Pro's 2K ceiling) | Seedream 4.5 |
Because every route shares one EvoLink endpoint and API key, this gate costs you one string: the
model field. Wire it as a config value, not a constant, and switching or falling back later is a deploy-free change.Step 2: Prepare Reference Images Like Inputs, Not Attachments
The single biggest quality lever in multi-reference generation is not the prompt — it is what you put in
image_urls and why. Treat the array as a structured input with roles:- Subject references — the product, person, or character that must stay consistent. Use clean, well-lit shots; the model can only preserve what it can see.
- Material/detail references — close-ups of texture, fabric, logo placement, packaging detail.
- Style references — an image whose lighting, grade, or composition you want transferred.
- Background/scene references — the environment the subject should be placed into.
- Brand asset references — layouts or visual-language samples that anchor the output to your brand system.
Practical limits to respect:
- 10 images maximum. If you are tempted to send more, your roles are probably muddled — cut style references before subject references.
- Order matters for your prompt. You will refer to images by position ("image 1", "image 2"), so keep a stable convention: subject first, then materials, then style, then background.
- Cost note: images 1-10 bill at 0.1734 credits each. A full 10-reference request adds 1.734 credits of input cost on top of the output tier — plan for it, but do not let it push you into sending fewer subject references. Cut decorative references instead.
Step 3: Pick Size and Output Format Safely
Tier choice. Use
1K for social content, drafts, and anything viewed on a phone. Use 2K when the image will be seen large — hero banners, print-adjacent uses, marketplace zoom views.Exact dimensions. Pro accepts
WIDTHxHEIGHT between 1024x1024 and 2048x2048 across ratio presets (1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3). Generating at the placement size beats generating square and cropping — you keep composition control and skip a pipeline step.The 2K boundary trap. Billing tier is decided by total output pixels: anything above 2,360,000 pixels bills as 2K. A 1536x1536 request (2,359,296 pixels) is 1K; 1600x1500 (2,400,000) is 2K. If your product exposes free-form dimensions, compute
width × height and show the tier before the user generates.Format. Default
jpeg for ordinary visuals — smaller files, universally displayable. Choose png when downstream work needs transparency or lossless input (compositing, print prep, further editing). Do not promise 4K anywhere in your product copy: the route's ceiling today is 2048x2048.Step 4: Build a Prompt That Tells Pro How to Use Each Reference
With multiple references, an unstructured prompt makes the model guess which image plays which role. Remove the guessing — assign roles by position:
Use image 1 as the product (keep shape, label, and colors exact).
Use image 2 for the fabric texture of the backdrop.
Use image 3 for lighting and mood only.
Place the product on a walnut shelf in a sunlit studio,
16:9 composition with space on the left for headline text.Patterns that hold up in production:
- One sentence per reference, role named explicitly ("keep exact", "style only", "background only"). This is the difference between style transfer and accidental subject replacement.
- State what must not change before what should. Preservation instructions ("keep the label text exactly as in image 1") outperform hoping.
- For edits, name the operation and the boundary: "Replace the jacket in image 1 with the jacket from image 2. Do not change the face, pose, or background."
- Layout hints belong in the prompt ("leave the right third clear for copy") since there is no layers control on this route.
Step 5: Estimate Cost Before You Batch
This section covers the mechanics — current per-tier prices live on the Seedream 5.0 Pro page and that page carries a verification date. The mechanics you should design around:
| Request shape | Output tier | Input images | Cost drivers |
|---|---|---|---|
| Prompt-only draft | 1K | 0 | Output tier only — the cheapest Pro request |
| Single-reference edit | 1K | 1 billed | Output tier + 0.1734 credits input |
| Full multi-reference (10 refs) | 2K | 10 billed | Output tier + 1.734 credits input — the most expensive shape |
| Batch of N variations | 1K × N | same refs each time | Input images bill per request, not per project |
The batching implication is the one teams miss: reference input cost repeats on every request. Ten variations from the same 10 references pay the reference surcharge ten times. If you are exploring compositions, do it at 1K with a trimmed reference set, and reserve the full reference stack for finals.
Step 6: Production Guardrails for Your UI and Backend
Before exposing Pro to end users or batch jobs, encode the route's contract as validation, not documentation:
- Cap reference uploads at 10 in the UI; reject at the API layer too.
- Fix
nto 1 and hide any batch-count control; implement variations as parallel requests. - Hide unsupported parameters — streaming, web search tools,
guidance_scale, layers,prompt_priority. Forwarding them invites errors or silent no-ops. - Show the billing tier pre-generation using the 2,360,000-pixel boundary.
- Treat generation as an async task: poll or take callbacks, persist the returned asset promptly, and design retry logic around failed tasks.
- Log the usage fields —
credits_used,input_images,billable_input_images,generated_images,output_tier,output_pixels— per request. They are your cost-attribution trail; there are no token counters to reconcile. - Wire a fallback route. If Pro degrades or a request needs sequential sets, fall back to Seedream 5.0 Lite by swapping the model string.
Example Workflows
Product re-dressing for e-commerce. Image 1: the product on white. Image 2: target fabric/colorway. Prompt: replace the material, keep geometry and label. Output at the marketplace's exact listing dimensions. Two billed input images add 0.34 credits of input cost per SKU variant.
Brand-consistent campaign series. Images 1-2: brand subject. Image 3: brand style sample. Generate each placement (1:1 feed, 9:16 story, 16:9 banner) as a separate request with the same reference stack, dimension per placement. Consistency comes from the shared references, not from a story-set feature.
Transparent PNG asset production. Prompt-only or single-reference requests with
output_format: "png", subject centered, plain background, sized for compositing. Feed the output to your existing cutout/compositing step.Localized ad visuals. Keep the reference stack fixed; vary only prompt language/copy-space instructions per market. Reference-locked subjects survive localization far better than re-prompting from scratch — the references are your consistency contract.
When Not to Use Seedream 5.0 Pro
- High-volume, cost-first drafting — Seedream 5.0 at $0.031/image is the batch route.
- Story sets / sequential images — not supported on Pro; use Seedream 5.0.
- Anything requiring live web context — web search tools are not available on this route.
- Deliverables promised above 2K — the ceiling is 2048x2048; do not sell 4K on this route.
- More than 10 references — Seedream 5.0 Lite accepts up to 14 if reference count is the constraint.
Rollout Checklist
- Route gate implemented (Pro vs 5.0 vs 4.5 as config, not constants)
- Reference upload capped at 10; roles documented for your content team
- Tier calculator wired to the 2,360,000-pixel boundary, shown pre-generation
- Unsupported params hidden and stripped server-side
- Async task handling: polling/callback, asset persistence, retry on failure
- Usage fields logged per request for cost attribution
- Fallback route tested with a one-line model swap
- Price display points at the canonical price table, not hardcoded numbers
- QA pass: single-ref edit, 10-ref generation, png output, exact-dimension request
FAQ
How many reference images can Seedream 5.0 Pro use at once?
Up to 10 per request via
image_urls. Each input image bills at 0.1734 credits.How do I keep a character or product consistent across many images?
Keep the same subject references in every request and assign them explicit roles in the prompt ("keep the product in image 1 exact"). Consistency comes from reference reuse, not from generating everything in one call.
Can I generate several images in one request?
No —
n is fixed to 1 on this route. Run parallel requests for variations; input images bill per request.When should I pick 2K over 1K?
When the asset is viewed large (hero banners, zoomable product views). Note the tier is decided by total pixels: above 2,360,000 pixels bills as 2K even for custom dimensions.
When should I output PNG instead of JPEG?
When downstream work needs transparency or lossless input — compositing, print preparation, or further editing. Otherwise jpeg is smaller and universally displayable.
Does Pro support story sets or sequential image generation?
No. Sequential sets are supported on Seedream 5.0 Lite, which is also the cheaper route.
Can it search the web or take live context?
No — web search tools are not available on this route. Provide all context via prompt and reference images.
Where is the current price?
On the Seedream 5.0 Pro page, which maintains the price table with a verification date. This guide intentionally documents mechanics rather than numbers so it does not go stale.
Is there streaming output?
No. The route is an asynchronous task: submit, then poll or receive a callback.
Sources: capability and parameter facts are verified against the EvoLink Seedream 5.0 Pro route and its product page (verified July 7, 2026); official parameter background for the Seedream image generation family is documented in the Volcengine image generation API reference. Where this guide and any third-party article disagree on route behavior, the EvoLink product page is the source of truth for the EvoLink route.
