Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.

Source too long? The 1800 s and 300 s caps for each Sume media tool
Trim, detach and inspect take sources up to 1800 s; filter, frames and compose stop at 300 s. The error each returns and the order to cut a long file.

Change one of three identical bottles: position words or a mask
Edit one of three identical products in a photo. Flux 3 Image gives each element an id; on Sume, use position words (Ideogram 4.5) or a GPT Image 2.5 mask.

Three-minute YouTube Short from six 30 s AI clips: Wan vs Seedance
YouTube counts square or vertical videos up to three minutes as Shorts. Six 30 s 9:16 clips cost $22.50 on Wan 3.0 at 720p or $104.04 on Seedance 2.5.

Caption design colors: hex, rgb(), rgba() or transparent only
Sume caption design colors take hex, rgb(), rgba() or transparent. Other CSS syntax is rejected at request time, so a bad color costs nothing.

30-second Wan hero plus two 15-second cutdowns: trim, not regenerate
One 30-second Wan 3.0 clip at 720p is $3.75. Two 15-second cutdowns by video trim add $0.04; regenerating them as 15-second clips costs $3.75 more.

Sume status vocab: a job is completed, a resource is ready
Sume lists three status vocabularies: job, resource, webhook delivery. Only jobs say completed; resources say ready, so a check on the wrong one never matches.

Volunteer recruitment video from an AI avatar: 25-second clip
A short avatar clip can ask for volunteers on a page, a newsletter or a social post. Script shape, three-scene option, caption settings and cost by tier.

Music Router prompt limit: 5,000 characters, and what to put in them
Sume Music Router prompts run 1 to 5,000 characters, with no duration field or negative prompt; length goes in the text. A brief budget and a checker.

SHA-256 the batch body into the Sume idempotency key for bulk chunks
Derive each bulk chunk's Idempotency-Key from a hash of its items so replays reuse the queue and edited rows get a new key instead of a 409 conflict.

GET result_url returns 409 run_not_completed: poll status_url first
A 409 run_not_completed from result_url means the Format run is not terminal. Read details.status, poll status_url with backoff, then fetch the result.

TypeScript union for Sume bulk queue items, checked with Deno
Model queue items by status so run_id and error are typed per case: null while queued, null for a child that never started, and an exhaustive switch.

Put a new picture on the TV or a frame in a room photo
Swap what is on a TV screen or in a picture frame in a room photo by sending the room first and your artwork second to Ideogram 4.5 on Sume. Cost and checks.

Remove a person or object from a finished clip with Gemini Omni edit
Delete a passer-by, a sign or a cable with one video_url request to Gemini Omni Flash 1.1. There is no mask field, so the prompt does the work. Cost per clip.

Chapter timestamps for narrated audio from concat segment offsets
Join one TTS file per chapter with timeline audio concat, then turn the returned segments[] start offsets into mm:ss chapter lines with a short Python script.

Alt text for a 30-image gallery in one Sume Agent Completion
One Agent Completion call can take up to 30 images and return an alts array under an object schema. Python stdlib script with the cap, poll and limits.

Add an object to a set spot in a photo: box to words in Python
Sume has no bounding-box field. Turn a 0-1000 box into placement words for an Ideogram 4.5 edit, and into pixels for a mask, with a short Python helper.

Investor update deck to a 15-second teaser: ten slides into Wan 3.0
Pick the ten slides that carry numbers and charts, send them to wan-3.0 as reference images, and ask for 15 s. $1.88 at 720p. Why charts need a real overlay.

MiniMax H3 or H3 Max after Sora: native 768p vs latent 1080p prices
On Sume, H3 renders native 480p or 768p and bills 2K/4K upscales; H3 Max adds 1080p as a latent refinement of 768p. 10-second prices side by side.

Ten-episode micro-drama on Wan 3.0: 30 s episodes, cost by tier
Ten 30-second Wan 3.0 episodes cost $18.80 at 480p, $37.50 at 720p and $75.00 at 1080p on Sume. Table, rounding check and a draft-then-final plan.

Replace the view outside a window in an interior photo
Swap the scene seen through a window and keep the curtains, frame and room, then relight in a second pass. Ideogram 4.5 on Sume, with prompts and costs.

Idempotency-Key from an order id and version, never a fresh uuid
A fresh uuid per request makes Idempotency-Key do nothing. Derive it from the order id plus a version you bump only to re-run. Scope: one Format, 255 chars.

Three characters, three dances: Omni IMAGE_REF and VIDEO_REF tokens
Google's Omni 1.1 demo swaps three dancers for a dog, an octopus and a bear. Here is the same request on Sume, with the 0-based reference tokens in order.

List every Format run: no GET /v1/format-runs, page per Format
GET /v1/format-runs does not exist. List runs per Format with limit, next_cursor and has_more, or keep your own index of the data.id you stored at create.

sume jobs cancel needs --confirm-submit, and only queued jobs cancel
Sume CLI cancel is a write: sume jobs cancel <job_id> --confirm-submit. It works only before generation starts; later: 409 job_generation_already_started.

Price per kept AI clip when 1 in 3 is rejected: Wan 720p
A 15-second Wan 3.0 clip at 720p is $1.875 on Sume. If you keep one clip in three, the real price per kept clip is $5.625. The keep rate is a budget input.

Which URL does each Sume media tool accept: public or media.sume.com?
Video captions and face swap take a public HTTPS URL; trim, filter, frames, inspect, detach, compose and timeline need your workspace's media.sume.com file.

Animate 12 storybook pages as 5-second clips: Omni or MiniMax H3
A picture book with 12 illustrated pages, each animated for 5 seconds. Price per book on Gemini Omni Flash 1.1 and MiniMax H3 on Sume, and the input each takes.

Does Wan 3.0 do 4K? Claim check against vendor pages and Sume
Alibaba lists 480p, 720p and 1080p for Wan 3.0, Morphic says no 4K, and Sume rejects 4k for wan-3.0. Where 4K exists on Sume and at what clip length.

Cancel a wrong video job after a Sora port: only before it starts
Ported prompts on the wrong model burn money. Sume cancels a video job only before generation starts; later you get 409 job_generation_already_started.

video-frames returned a null url: the job succeeded, one frame failed
If one instant fails to extract, video-frames sets that frame's url to null and the job still completes. How to detect it and retry only that time.

Crop a 2:1 landscape phone recording to 9:16: width 0.28125
A 2:1 (18:9) landscape clip needs width 0.28125 for 9:16: exact at 1920x960 (540x960), 606 px at 2160x1080. The offset, the rounding and the full request.

video-filter unsupported_filter_op_field: crop and dim keys allowed
A crop op takes four fractions and dim takes one amount; any extra key returns unsupported_filter_op_field. The exact key lists and where the extra work goes.

jq one-liner: export Sume bulk queue items to CSV by index
Turn the bulk queue receipt from a curl poll into CSV rows of index, status, run_id and error code with jq, then paste them next to your SKU column.

Set your Format run timeout from expires_at, not a guess
A non-terminal Format run receipt carries expires_at, the deadline after which Sume force-finalizes it as failed. Derive your wait from it and read queue.state.

Avatar first-frame preview checklist: seven checks before final render
Before paying for the full avatar render, check seven things on the preview stills: handle, framing, background, product, ratio, scenes, caption intent.

Scam alert video for customers with an AI avatar: script rules
If scammers use your name, a 20-second avatar clip can warn customers. Script rules, the AI label, why it cannot verify anything, and cost by tier.

Change the name and number on a jersey photo, keep the lettering
Edit the name and number on a jersey or sign photo with Ideogram 4.5 on Sume: the prompt wording, the reference rules and what a low-tier test costs.

Python preflight for a Sume bulk body: count, keys, bytes, caps
Check a Sume bulk-run body offline before the POST: 1 to 100 items, concurrency 1 to 16, 64 input keys, 2 MiB input, 4 MiB body and a spend cap in range.

Check an Omni edit's length with video-inspect before a timeline join
An edit should follow the source length. Confirm it with a probe-only video-inspect call before the clip goes into a Timeline render, with a Python read.

Check an Omni edit kept the rest of the clip: video-frames pairs
Compare stills from the source and the edited clip at the same timestamps with video-frames on Sume. A script that submits both extracts, plus what to look for.

Validate Wan 3.0 reference limits in Python before submitting
wan-3.0 takes 10 images, 5 videos (15 s total, 16 fps minimum) and 5 audios (15 s total). A short Python check catches an over-limit manifest early.

Four characters, two locations: a Wan 3.0 series bible in ten images
Wan 3.0 on Sume takes 10 reference images per job. Spend them as four characters at two views each plus two sets, and reuse the same set for every episode.

Photo to talking selfie clip: put the spoken line in the Omni prompt
fal's Omni 1.1 example ends its prompt with a quoted spoken line. Here is that pattern on Sume from a still, the cost of 6 seconds, and a transcript check.

Destination film from ten location photos with Wan 3.0 on Sume
Turn ten photos of a town into a 30-second film with wan-3.0 reference images: $3.75 at 720p, $7.50 at 1080p, and a prompt that needs a dawn-to-dusk arc.

Reuse one sentence twice: timeline audio split allows overlaps
Timeline audio split takes up to 20 ranges that may overlap, so one hook sentence can open and close a video without re-recording or re-synthesising it.

A 2.5% word error rate is 25 wrong words per 1,000: plan the proofread
If a 2.5% WER held on your audio, a 1,000-word transcript would still carry about 25 wrong words. How to proofread and burn captions on Sume with script_text.

gemini-3.1-flash-image is deprecated: move to Nano Banana 2.1
Google deprecated gemini-3.1-flash-image on Oct 6 with no shutdown date and named gemini-nano-banana-2.1 as the replacement. What to change, and what Sume does.

Nano Banana Pro vs 2.1: 4x the price at 1K, 2.1x at 4K
Google lists Nano Banana Pro at $0.134 per 1K or 2K image and $0.24 at 4K, against $0.0336, $0.0504 and $0.113 for Nano Banana 2.1. When Pro still pays.