Direct answer
The short version
Treat media input as a versioned API contract, not a generic upload field. For every provider, model and region, record the accepted URL or Base64 form, file type, byte size, pixel dimensions, aspect ratio, number and order of reference files, request-body limit and how long a temporary URL must remain reachable. Run a small synthetic preflight set through task creation, processing and result retrieval before exposing the workflow to customer media.
Why can the same image pass one video API and fail another?
Chinese AI video APIs expose different media contracts even when they all advertise image-to-video. Alibaba Cloud's Wan reference currently accepts a public URL, an Alibaba OSS temporary URL or a Base64 data URI, while its allowed formats, file-size ceilings and pixel range vary by model generation. MiniMax's current image-to-video endpoint accepts a public URL or Base64 data URL but documents its own file-size, short-edge and aspect-ratio rules.
A provider can also enforce a request-body ceiling that differs from the media-file limit. Tencent Cloud's current TokenHub Vidu guide, for example, documents URL or Base64 image input together with both a media limit and a smaller POST-body limit. Validate the complete request envelope instead of checking only the source file on disk.
- Provider, account, region, endpoint and exact model ID
- Input mode: first frame, first and last frame, subject reference or multiple references
- Transfer form: public URL, provider temporary URL or Base64 data URI
- Media constraints and total JSON request-body constraint
- The provider page and access date used for approval
Which file properties belong in a preflight check?
Check the binary file rather than trusting its filename. Confirm the decoded format, byte size, width, height, orientation and alpha-channel behaviour before submission. Normalize metadata rotation where necessary, because a portrait image whose pixels remain landscape can produce a different aspect ratio after decoding.
Keep limits tied to the model version. Alibaba's current Wan documentation distinguishes newer and older model-series limits, and MiniMax lists its requirements on the selected image-to-video task endpoint. A single global rule copied into application code will drift when a provider changes a model or when a team adds another route.
- Decoded MIME type and extension agreement
- File bytes before and after any Base64 encoding
- Pixel width, height, short edge and aspect ratio
- Alpha channel, colour profile and orientation metadata
- Number, role and order of reference files
How should a public or signed media URL be tested?
The provider's fetch service, not the buyer's browser, must be able to reach the URL. Test from a neutral environment and require a direct successful response without a login page, expiring session cookie or private network dependency. Confirm the returned content type and bytes, not only an HTTP success code, because an error document served with status 200 is not a valid image.
A signed URL needs enough lifetime for queue delay, provider fetch, bounded retries and investigation. Do not set its expiry from the expected generation time alone. Record whether redirects are accepted, whether the storage host blocks unfamiliar user agents, and whether a refreshed signature changes the job's idempotency decision.
- HTTPS reachability without interactive authentication
- Correct content type, content length and binary signature
- Signed-link validity beyond the tested processing window
- Redirect, firewall, hotlink and geographic access behaviour
- Revocation and cleanup after the bounded test
When should Base64 be used instead of a URL?
Base64 can remove a separate public hosting step, but it increases the request size and can expose media to application logs, error traces or queue payloads. Use it only when the provider and chosen model document support for the exact data-URI form and the encoded request stays below every gateway and provider body limit.
Keep Base64 generation inside the customer's controlled environment. Do not paste production media into consoles, tickets or diagnostic reports. Log only a content-free identifier, original byte count, decoded dimensions, MIME type and a one-way checksum when operationally necessary.
- Generate a complete data URI with the documented MIME type
- Measure the encoded payload and full request body before sending
- Exclude request bodies from routine logs and error monitoring
- Cap retries so a large payload is not resent without control
- Prefer synthetic media for integration diagnostics
How do reference modes change the input contract?
First-frame, first-and-last-frame and subject-reference modes are separate contracts. MiniMax documents distinct video-generation modes, while Tencent's Hy guide says image-to-video is selected by supplying an image or image URL and that the input image determines the output ratio for that mode. Do not assume fields, model availability or output controls transfer between modes.
Create one test case per mode actually required by the workflow. Verify which image controls the frame ratio, whether reference order is significant, how many files are accepted, and what happens when optional prompt or audio fields are combined with media input.
- One first-frame image with a supported ratio
- First and last frames with compatible dimensions where supported
- Subject reference using non-confidential synthetic material
- Boundary cases for supported reference count and order
- A deliberate invalid case for each preflight rule
What should the pre-production test matrix contain?
Start with a small, versioned set that contains normal files plus one boundary and one invalid example for each rule. Stop after the planned attempts. A rejection before task creation, a provider moderation failure and a generation failure are different outcomes and should remain separate in the record.
Run the media through the whole asynchronous path. Successful task creation does not prove the provider fetched every reference or that the result can be retrieved. Record sanitized provider request IDs and errors, but do not retain prompts or media in the operating report.
- URL and Base64 versions of the same synthetic image
- Minimum, normal and near-limit file dimensions and sizes
- Landscape, portrait and square aspect ratios used by the business
- Temporary-link lifetime longer than the measured queue and processing window
- Task acceptance, terminal state, output download and creative review
Which decision criteria make an input path ready?
Approve only the combinations that pass a written matrix for the exact account, region, endpoint and model. The record should distinguish provider-documented limits from limits observed in the test. If the live behaviour conflicts with the documentation, stop and resolve the discrepancy before increasing volume.
IT CaoCao can prepare a fixed-scope input matrix and review the resulting evidence without receiving production credentials or customer media. The customer controls storage, signed URLs, provider accounts, billing and final retention decisions.
- All required business formats pass preflight and provider processing
- Temporary URLs remain reachable for the full tested lifecycle
- Logs and support records exclude media and request bodies
- Failures map to actionable validation messages without automatic resubmission
- Owner, source date and retest triggers are documented
Frequently asked questions
Questions buyers ask next
Is a browser-accessible image URL enough for an AI video API?
No. The provider's fetch service must reach the URL without interactive authentication and receive the expected media bytes. Test content type, redirects, signed-link lifetime and storage access controls.
Can every Chinese AI video API accept Base64 images?
Do not assume so. Verify the exact provider, model and field documentation, the required data-URI format and the encoded request-body limit before enabling Base64 input.
Should an application copy one provider's image limits to every model?
No. Store validation rules by provider, region, endpoint and model version because formats, dimensions, aspect ratios and byte limits can differ and can change over time.
What media should be used for integration testing?
Use approved synthetic or licensed non-confidential files that cover normal, boundary and invalid cases. Keep production customer media, credentials and full request bodies out of test reports.
Primary sources
Evidence reviewed
Provider documentation changes. Recheck the live source before making a procurement decision.
- Alibaba Cloud Model Studio: Wan image-to-video first-frame API referenceAccessed 2026-09-07
- MiniMax API: image-to-video taskAccessed 2026-09-07
- Tencent Cloud TokenHub: Vidu image-to-video guideAccessed 2026-09-07
- Tencent Cloud TokenHub: Hy video generation guideAccessed 2026-09-07