Seedream image jobs now log whether image input was present, whether it was sent as a single string or an array, how many references were included, and which size preset was used. This makes it possible to prove whether missing reference images are a server bug or a downstream gateway issue.
Constraint: The test server needed a low-noise way to confirm what reached /v1/images/generations without changing the external API contract
Rejected: Rely on gateway-side observation alone | it could not distinguish whether the image array was dropped before or after the server boundary
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep this Seedream request-shape log until the gateway team has finished validating image-array forwarding end to end
Tested: go test ./internal/server
Tested: Seedream 4.5 job on test server with three image refs after deploying binary (job_47513ead2708)
Not-tested: Long-term log volume impact under sustained Seedream traffic
The server now resolves source, identity, and style artifacts separately, emits role-labeled Gemini parts, and logs the effective request shape so deployed debugging can confirm which references reached upstream.
Constraint: The test deployment must stay on gemini-3.1-flash-image-preview while gaining better multi-image editing semantics
Rejected: Change the default model to gemini-3-pro-image-preview | test-server parity required keeping flash as default
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Preserve the source/identity/style logging until the upstream gateway exposes equivalent structured traces
Tested: go test ./internal/server
Tested: image img2img against http://101.42.99.35:18080/v1 (job_5af212df2c27)
Not-tested: Non-Gemini multi-image providers still fall back to the first reference only
The server now keeps Vidu URL-based image2video requests on a valid default duration when callers omit duration, and it forwards MiniMax S2V reference images in the images[] shape expected by the deployed Hailuo adapter.
Constraint: The test environment currently validates provider behavior through popiartServer binary deploys rather than waiting for broader upstream adapter rollouts
Rejected: Keep the old shared 4-second default for all providers | breaks viduq3-turbo when duration is omitted
Rejected: Translate S2V references into subject_reference in popiartServer | the deployed Hailuo adapter already expects top-level images and performs its own mapping
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep provider-specific video defaults and reference-shape compatibility isolated in popiartServer unless upstream adapters are upgraded in lockstep
Tested: go test ./internal/server -run TestInferRouteKeyForModelRecognizesViduAsVideo|TestGenerate|TestResolve|TestSubmitMiniMaxVideoTask|TestSubmitImageToVideoTask|TestExecuteImageToImageJobUsesGenerationsPathForMiniMax|TestExecuteImageToVideoJob|TestResolveVideoReferencesSupportsImagesArray
Not-tested: Full PopiNewAPI unit suite from the local checkout
The server now keeps portable image ratio normalization, enforces documented Seedream size presets, routes MiniMax img2img through generations-style payloads, and submits Hailuo video variants through the video generations API with proper reference-image handling.
Constraint: The deployed test chain relies on popiartServer as the lowest-risk place to adapt provider-specific routing without broad newapi changes
Rejected: Expand newapi relay modes for every MiniMax image/video branch | larger blast radius than needed for the verified server-managed flows
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep MiniMax-specific media routing in popiartServer unless upstream newapi adapters are upgraded to the same contract
Tested: go test ./internal/server -run TestInferRouteKeyForModelRecognizesViduAsVideo|TestGenerate|TestResolve|TestSubmitMiniMaxVideoTask|TestExecuteImageToImageJobUsesGenerationsPathForMiniMax|TestExecuteImageToVideoJob|TestResolveVideoReferencesSupportsImagesArray
Not-tested: Full PopiNewAPI unit suite (blocked earlier by external module checksum drift in local checkout)
Seedream image generation in the server was still split across the generic text2image path and the legacy multipart image edit path. That shape does not match Volcengine Seedream 4.5, which expects JSON requests to /v1/images/generations for both text2image and img2img, with reference images carried through the image field.
Constraint: Seedream 4.5 and 5.0 use /v1/images/generations rather than the older multipart /v1/images/edits contract
Constraint: Existing Gemini and other image routes must keep their current behavior unchanged
Rejected: Keep using /v1/images/edits for Seedream img2img | upstream rejects the request body shape
Rejected: Special-case the CLI input contract | the server route adapter owns upstream normalization
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep Seedream routes on the JSON generations path; do not reintroduce multipart edits for these models without revalidating the upstream API contract
Tested: go test ./internal/server; go test ./...; go build -o /tmp/popiartserver-verify ./cmd/popiartserver; test-server txt2img/img2img/img2video smoke with Seedream 4.5 + Vidu
Not-tested: Seedream 5.0 live smoke against the test server
jiajia
张正强