Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
Cancel one episode in a running Sume bulk queue: what it frees
There is no cancel-queue endpoint on Sume. Cancel one item run and its slot frees for the next episode. A canceled run delivers no webhook; read cancel_effect.
YouTube suggests 3 Shorts thumbnail frames: pick from 24 yourself
YouTube suggests three frames for a Shorts thumbnail. Pull up to 24 candidates with Sume video frames and choose your own.
Black first frame in a Short? Fade-in and your thumbnail frame pick
output.fade_in_seconds in Timeline 1.0 starts the video from black, so frame 0 is a poor thumbnail. Pick a frame after the fade with Sume video frames.
Reels ad: use an existing post or upload a new video? Sume 9:16 file
Meta's Reels ads page offers native video, image ads and existing Reels posts. Which route fits an AI video from Sume, and what file to prepare for each.
Reels image ad with auto music, or your own video? Still to 9:16
Meta's Advantage+ Creative can add music to Reels image ads. When to instead turn the still into a 9:16 video with image-to-video on Sume.
TikTok Symphony Agent adds AI labels; a Sume upload does not
TikTok says Symphony Agent output carries AI labels and invisible watermarks. A Sume file has no such step in the docs, so here is what to add before posting.
LinkedIn video ad needs an uploaded file, not a link: use the Sume MP4
LinkedIn's video ad spec says to upload the file, not link YouTube or Vimeo. How to take the MP4 from a Sume job, check it, and upload it.
Poll a Sume job from a Node script: deadline, backoff and exit code
A short Node script polls a Sume job with getApiJobStatus, honors next_poll_after_seconds, and gives up at a deadline without cancelling the job.
Render a Short and wait for it in one request? Timeline sync mode
Timeline 1.0 can answer a Short render in one request when mode is sync and the job finishes within 30 seconds. Otherwise it returns 202 and you poll.
Qwen Image vs Qwen Image Max on Sume: edit support and a 3.75x price
Qwen Image bills $0.025 and edits photos; Qwen Image Max bills $0.09375 and is text-to-image only on Sume. Both list 13 ratios, including 21:9 and 9:21.
Read the Sume TTS Router catalog in one request: Sonic ids, prices
GET /v1/tts-router/models lists the Sonic ids, the per-character list price, the margin and the 20,000-character cap. A small Node script prints them.
Codex 0.160.0 resumes queued messages after reconnect: Sume key rule
Codex 0.160.0 resumes unsent queued messages after a reconnect without duplicates. A paid Sume call still needs its own idempotency_key to be safe.
Microdrama TV views are up 90%: render episodes at 1080p?
YouTube says microdrama views on TV rose over 90% year over year. Episodes stay vertical, but 1080p on Wan 3.0 costs $0.25 a second, double 720p.
Recurring cast in Gemini Omni Flash 1.1: 10 images, 3 short clips
Omni Flash 1.1 reference-to-video on Sume takes up to 10 images plus 3 clips of 3 s or less, addressed as IMAGE_REF_0. A 10 s 9:16 shot at 720p is $1.25.
Previously on: add a recap to a YouTube Shorts series episode
Open a Shorts episode with a Previously on recap: build it with Timeline 1.0 audio.parts and see what it costs per minute.
YouTube thumbnails take JPG or PNG: output formats per Sume image row
YouTube accepts JPG and PNG thumbnails. Sume rows list png, jpeg or webp; Recraft V4 is webp only, Soul and Ideogram 4.5 pick the format for you. Table and fix.
Claude Code 2.1.293 HTTP MCP leak fix and long Sume job sessions
Claude Code 2.1.293 fixed a leak where an HTTP MCP connection kept every request it sent. Upgrade before a long Sume session and use batch jobs_wait.
How many new Sume jobs can I submit now? Read generation_limits
Open slots are concurrency_limit minus active and queued jobs, capped by queue_capacity_remaining. A short Python function computes it from a submit response.
Submit 12 thumbnail jobs with mode async and Idempotency-Key on Sume
Queue a dozen 16:9 thumbnails on POST /v1/images with mode async and one Idempotency-Key each. Re-run the script safely, then poll the status URLs. Python code.
A/B test three hook lines on one Short with caption cues
YouTube says A/B testing of up to three video cuts is coming soon. Prepare three hook-line variants from one clean Short with Sume caption cues at $0.20 each.
Seedance 2.5 vs Kling 3 on Sume: what each catalog row accepts
Seedance 2.5 takes 4 to 30 seconds, eight ratio values and image, video and audio references. Kling 3 takes 4 to 15 seconds, three ratios and no references.
Claude Code 2.1.287 tool heartbeats and Sume jobs_wait slices
Claude Code 2.1.287 fixed tool heartbeats not reaching SDK hosts during a stalled response stream. Sume's jobs_wait holds up to 55 s, so heartbeats matter.
Shorts series sequential playback: fade out or hard cut each episode?
YouTube plays Shorts series episodes in order. Pick a hard cut or a short fade at the end of each one, with the Timeline 1.0 fade limits.
Binge cut: join ten AI micro-drama episodes into one video
Ten one-minute episodes become one 10-minute video in a single Timeline 1.0 render: detach each episode's audio, join the parts, and expect a $1.00 render fee.
Image 1.0 to POST /v1/images: image_urls, num_images, mask mapped
Moving from /v1/image-1.0/generate to /v1/images on Sume: image_urls becomes input_references, num_images becomes n, and the default mode becomes sync.
Gemini voice replication: the 10-30 second clip and consent checklist
Gemini 3.8 voice replication needs a 10-30 second clean clip and a spoken consent statement. A checklist, and why a Gemini voice name will not work on Sume.
Shorts series: serial or non-serial for AI episodes
Shorts series can be serial (oldest first) or non-serial (newest first). Which suits AI episodes, and how Sume jobs differ for each, per YouTube Help.
Three-second bumper: Gemini Omni Flash 1.1 or Wan 3.0 on Sume
A 3-second bumper costs $0.38 at 720p on Omni Flash 1.1 and $0.38 on Wan 3.0. Omni starts at 3 seconds; Wan starts at 2.
Same set every episode: an Avatar 1.0 scene photo or prompt background
Avatar Video 1.0 scenes are a prompt or a photo. Use one photo for the set in every episode, or reuse one prompt word for word, to keep the same place.
script_run or bulk runs: where to fan out 20 ad variants
script_run fans out three or more Sume tool calls inside one 55-second request; bulk runs queue up to 100 Format runs for hours. Limits and a rule of thumb.
A small 540x960 draft of a Short for quick feedback: output size
Render a lighter 540x960 draft of a Short with Timeline 1.0 output width and height. The size is free to choose; the price per minute is the same.
LTX-2.5 Fast 20-second clip: Sume's nearest rows with Wan 3.0 prices
LTX-2.5 Fast runs up to 20 seconds. Sume has no LTX id; Wan 3.0 covers 20 seconds for $1.25, $2.50 or $5.00.
Claude Code 529 retry delay setting is not a Sume 429/503 retry
Claude Code 2.1.290 added a base delay setting for 529 retries. It does not govern Sume calls; follow retry-after and keep the Idempotency-Key.
MiniMax H3 Max API in Python: reference-to-video with three images
Use MiniMax H3 Max on Sume from Python: send three product photos in input_references, ask for 768p, poll and save. Reference limits and 10-second price.
Claude Code MCP tool name over 128 characters: does Sume hit it?
Claude Code 2.1.292 leaves out an MCP tool whose name passes 128 characters and names it in an error. Sume's longest tool name is 36, so it never trips it.
Continue a scene in the next Shorts episode: last frame as first frame
Pull the last frame of episode 1 with Sume video frames and use it as the first frame of an Omni clip for episode 2. Sequence, API calls and an 8-second price.
Kling 3 API in Node: a silent clip with generate_audio false
Call Kling 3 on Sume from Node: send generate_audio false for the cheaper silent rate, add a last frame, poll and save. A 10-second silent clip costs $1.40.
Exact 9:16 Shorts thumbnail with gpt-image-2.5: use 1152x2048
gpt-image-2.5 on Sume takes custom sizes in multiples of 16. 1152x2048 is exactly 9:16 and above the pixel minimum, while 576x1024 is below it.
Which Sume image models accept quality, and which reject it with a 400
Only four Sume image families list a quality value: ChatGPT Image 2.5, ChatGPT Image 2, Ideogram V3 and Ideogram 4.5. On any other row quality is a 400.
One caption per carousel slide: get a typed list from a Sume Format
Write a separate caption for each carousel slide by binding an output_schema to a Sume Format run, so your app gets one typed caption per image.
Sample clip for a YouTube brand pitch: a 6-second Omni preview
Make a short sample clip to pitch a brand on YouTube: Omni 1.1 prices per second and how to hand it over with Ask Studio's pitch help.
Codex 0.161.0 /mcp login: sign in to Sume without leaving the session
Codex 0.161.0 added /mcp login <name> to sign in to an MCP server from a live terminal session. Sign in to Sume with read access first, then add write.
VS Code 1.141 shows discovered MCP servers: where is the Sume entry?
VS Code 1.141 (Oct 7, 2026) lists MCP servers found outside your mcp.json in the Customizations editor. Check there for Sume's URL and the credential it uses.
Wan 3.0 API in Node: vertical image-to-video from a first frame
Call Wan 3.0 on Sume from Node with fetch: pin a first frame, ask for 9:16 at 720p, poll the job and save the MP4. Includes limits and the 8-second price.
What n: 4 costs on each Sume image model: cost x n, billed table
Sume's image docs say you pay cost_usd x n. Four images cost 4x the per-image rate: $0.10 on Qwen Image, $0.40 on Nano Banana 2.1, $0.75 on Pro. Full table.
Seedance 2.5, Seedance 2 or fast: pick the model for episode length
On Sume, Seedance 2.5 does 4-30 s, Seedance 2 does 4-15 s, and fast and mini top out at 720p. How the episode length picks the model, with the catalog call.
Grok Imagine Video 1.5 for episode teasers: key art to 4-15 s
Sume lists Grok Imagine Video 1.5 as image-to-video only, 480p or 720p, 4-15 s. Use it to animate one key-art still into a teaser, not to cast a scene.
A zod schema for the Sume /v1/videos poll response, ported from Sora
Replace Sora's four-status type with a zod schema for Sume's five statuses. It rejects a completed job without unsigned_urls and a failed one without an error.