Guides
Sume guides: the API, Formats, agents, models, media tools and pricing, each answer dated and sourced.
Veo 3.1 reference images: three max, and the 8-second rule
Veo 3.1 takes up to three reference images, and a request with them must run 8 seconds. Sume does not list Veo; its reference-capable models take more.
Veo 3.1 vs Kling 3.0: how to choose by inputs and limits
Google's Veo 3.1 limits next to the kling-3 entry in Sume's catalog: clip length, resolution, references and audio. Sume lists kling-3 and does not list Veo.
Veo 4 release date: what Google's Veo pages list today
Google's Gemini API pages list Veo 3.1 and its Lite preview and name Gemini Omni Flash the default video model. No Veo 4 date appears; here is what does.
Vercel AI SDK tool approval: confirm a paid video call
In the AI SDK, set toolApproval to 'user-approval' on a tool that starts a paid video job. The call pauses until your app sends back a decision.
verifyWebhook async: forget await and you get a promise
verifyWebhook in @sume-com/sdk is async. Without await, a Promise is truthy, so your signature check never rejects. The await, raw body and replay rules.
Microdrama script format: cliffhanger and AI shot list
Write a microdrama as a beat sheet, then turn each beat into one AI shot no longer than the model allows. A format with the cliffhanger as the last frame.
Crop a video by fractions and the out-of-bounds error
The video filter crop op takes fractions of the frame, not pixels. Rules, the 0.05 minimum, the out-of-bounds error, and a centered 9:16 crop of a 16:9 clip.
Darken a video with an API: the dim operation and its amount
The dim op multiplies a clip's brightness by amount in (0, 1]. 0.45 darkens, 1 changes nothing, 0 and above 1 are refused. Request, limits and cost.
seek fast or precise: which to use for video inspect stills
Video inspect's seek option: precise decodes the exact instant, fast snaps to the keyframe at or before it and skips the decode. What changes in the result.
Transcribe a video into sentence segments with Sume: options and cost
video inspect with transcribe true returns text, words and, with sentence segmentation, gapless caption-line segments. Options, the $0.01 per minute rate.
Choosing an AI video API as a developer: a technical checklist
Choose an AI video API on job model, not demos: async submit and poll, idempotent retries, webhooks, live capability discovery, a billing model you can read
AI video with audio: which models always add sound, which are optional
On Sume, minimax-h3, minimax-h3-max and gemini-omni-flash-1.1 always add native audio; kling-3, Seedance and Wan make it optional; grok-imagine-video-1.5 none.
Which AI video model should I use? A checklist by input and output
Pick an AI video model by what you must send and get back: frames, references, audio, length, resolution. This checklist maps each need to Sume catalog ids.
unsupported_media_source: why the media API rejects your video URL
unsupported_media_source means video_url is not on the Sume media host. Import the clip first; which Sume video endpoints need a hosted URL and which don't.
video_trim_range_conflict: send end or duration, not both
The video trim API returns video_trim_range_conflict when a body has both end and duration. Send start plus exactly one of them; other range errors explained.
Wan 3.0 open source: weights on Hugging Face or API only?
fal says Alibaba's published Wan weights stop at Wan 2.2, and Wan 3 is an API model. What that means for Hugging Face searches, and the wan-3.0 id Sume hosts.
Wan 3.0 prompt guide for API users: prompt vs request fields
Write a Wan 3.0 prompt as a shot list with sound cues, and set length and size in request fields. What to put in the prompt and what to send to Sume's wan-3.0.
Wan 3.0 2-second clips: the shortest video length and what it costs
Wan 3.0's floor is 2 seconds, shorter than the 4-second minimum on Seedance 2.x models. Send duration 2 to wan-3.0; at 720p it costs $0.25 before the agent fee.
Wan 3.0 document to video: what the model reads and what Sume passes
Alibaba's Wan 3.0 can read an uploaded document to make a video. On Sume's video API today wan-3.0 takes prompts, frames and media references, not a file input.
Wan 3.0 first and last frame: animate between two images
Wan 3.0 can make a clip that starts on one image and ends on another. On Sume, send frame_images with first_frame and last_frame to wan-3.0; last needs first.
Wan 3.0 omni reference video: images, clips and audio in one prompt
Omni reference means one request carries mixed references. Wan 3.0 takes up to 10 images, 5 videos and 5 audio tracks; on Sume, send them as input_references.
Wan 3.0 pricing: what a clip costs per second at 480p, 720p and 1080p
Wan 3.0 is priced per output second by resolution. fal lists $0.05, $0.10 and $0.20 a second; Sume bills that list price × 1.25. Costs for 2 to 30 s clips.
Wan 3.0 reference limits: how many images, videos and audio clips
Wan 3.0 reference-to-video takes up to 10 images, 5 video clips and 5 audio tracks. See the per-type caps, what Sume rejects with unsupported_capability, why.
Wan 3.0 seed and size: why the API returns 400 unsupported_parameter
wan-3.0 on Sume rejects seed and size rather than ignoring them. Use resolution and aspect_ratio instead. Why the 400 happens and how to fix the request.
Wan 3.0 thinking mode: why document and web page input needs it
fal says Wan 3.0 needs thinking mode to read a document or web page; Alibaba says prompt_extend must be true. What each means, and why Sume's wan-3.0 skips it.
Wan 3.0 vertical video: 9:16 for Reels, Shorts and TikTok
Wan 3.0 supports 9:16. Send aspect_ratio 9:16 to wan-3.0 for a vertical clip of 2 to 30 seconds at 480p, 720p or 1080p. Accepted ratios and cost.
Wan 3.0 vs MiniMax H3: clip length, ratios and reference caps
Wan 3.0 runs 2 to 30 s at up to 1080p; MiniMax H3 runs 5 to 15 s at 480p or 768p. Where they differ on Sume: ratios, reference caps and audio inputs.
Wan 3.0 vs Seedance 2.0: which model to call when
Wan 3.0 and Seedance 2.0 are both on Sume's video API. Wan runs 2 to 30 s and bills per second; Seedance 2.0 runs 4 to 15 s with a 21:9 frame. Pick by the job.
Wan 3.0 web page to video: link input and how to do it on Sume
Wan 3.0 can read a public web link to make a video on Alibaba's API. Sume's wan-3.0 has no link field, so fetch the page yourself and send text and images.
Webhook not fired when a job is canceled? Jobs vs runs in Sume
A canceled Sume job sends job.canceled, but a canceled run sends no webhook at all. Which surface you called decides it; how to detect a cancel by polling.
Webhook payload too large (1 MiB)? Fetch the result URL
A Sume run receipt over 1 MiB arrives with payload null and a payload_too_large error. The run did not fail; fetch the receipt from result_url. Handler steps.
Webhook endpoint redirect 301 not followed: what Sume does
Sume does not follow redirects on webhook deliveries: a 3xx is a failed attempt, not a delivery. Point webhook_url at the final URL, not a hop.
Webhook dedupe id stable across retries: request_id and created_at
On Sume run webhooks, dedupe on the envelope request_id, which equals run_id and repeats on every retry. Order by created_at; request_id cannot.
Webhook retry: fixed delay vs exponential backoff, with numbers
Fixed delay retries at even gaps; exponential backoff doubles them. Sume uses fixed for job webhooks and backoff for run webhooks: how long each keeps trying.
Webhook unknown event type: return 204, not a 500
One Sume verifier covers run and job webhooks. Route on event and answer 204 for unknown types so a new event never becomes a 500 and a retry storm.
Webhook signature mismatch: compare the secret fingerprint first
When a Sume webhook signature will not verify, compare x-sume-webhook-secret-fingerprint with the dashboard fingerprint. It is safe to paste into a ticket.
Webhook secret rotation: no downtime with a 24-hour window
After you rotate a Sume webhook signing secret, deliveries carry two signatures for 24 hours. What the header looks like and how to redeploy safely.
Webhook send test vs redeliver: which one replays a real job?
Send test posts a dummy webhook.test payload to a URL you type. Redeliver re-sends a real terminal event and does not use one of the automatic 10 attempts.
Webhook status OK but output null? Read outcome: degraded
A Sume run webhook can say status OK while output is null. The run completed and billed; outcome is degraded and output_error says why. How to branch on it.
What counts as significantly edited by AI on TikTok ads?
TikTok Ads says 'significantly modified by AI' means changes beyond minor tweaks. Lighting, background and denoising edits are insignificant.
What does 'Captured with a camera' mean on YouTube?
It is a YouTube disclosure for video from C2PA-capable tools whose sound and visuals were not edited. It is separate from the AI-content labels.
Why is TikTok labeling my video as AI-generated?
TikTok Shop's 09/02/2026 policy says the platform may auto-label content with specific technical metadata. Creators cannot remove that label.
Word document to video AI: turn a .docx into a narrated video
To turn a Word document into a video, extract its text and images first. Sume's agent takes text in input and images as attachments; a .docx isn't one.
WordPress cron not working: schedule AI video jobs reliably
WP-Cron runs only when someone loads a page. Trigger wp-cron.php from a system cron so scheduled AI video jobs start on time.
WordPress webhook endpoint: receive Sume video callbacks
Register a REST route with register_rest_route, verify the signature on the raw body in PHP, and answer 2xx so Sume stops retrying.
wp_remote_post timeout: calling a video API from WordPress
wp_remote_post waits 5 seconds by default. Set timeout yourself, submit the video job, and let a callback deliver the result instead of waiting.
YouTube AI music disclosure: when background music needs a label
YouTube's disclosure page lists AI generated music among content to disclose when realistic. What it says, what is exempt, and how a Sume music file fits in.
YouTube altered or synthetic content disclosure: what counts
YouTube asks for disclosure only when altered or synthetic content looks realistic. Sort your own edit steps into needs it, exempt, and not listed.