Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
HOA or community meeting notice video with an AI avatar: script, cost
A homeowners' association can post a 30-second avatar notice for the monthly meeting. What to say, 16:9 versus 9:16, review steps, and the cost by tier.
Do fades shift my cuts? Timeline start times with transitions
In Timeline 1.0 a slot's declared start is its place on the audio spine, fade or no fade. The compiler compensates for the xfade overlap; gaps hold a frame.
SQLite ledger for Sume bulk queues: SKU to run id and what to retry
Record each bulk queue item by index in SQLite, keep the SKU you sent, and query the SKUs that failed or were canceled and never completed in a later queue.
Go client for a Sume bulk queue: transient polls and exit status
Create a Sume bulk queue from Go, poll the status_url with a doubling gap, treat 429 and 503 as transient, and exit 1 on failed items. Standard library only.
URL or QR code in an AI avatar video: say it, caption it, or link it
An avatar clip cannot reliably carry a QR code or a long URL. Three Sume-supported options: spoken words, authored caption cues, or the text beside the video.
Pilot STT in 3 languages: language_code hint vs auto-detect, 30 clips
MAI-Transcribe-2-Streaming advertises 60 languages; here is a 30-clip pilot on Sume STT that compares a language_code hint with auto-detect for about 60 cents.
Crop a 5:4 monitor clip to 9:16: width 0.45 and the exact sizes
A 5:4 clip crops to 9:16 with width 0.45 and x 0.275. At 1280x1024 that is 576x1024, at 1600x1280 it is 720x1280. Two sizes, one body, and why it is exact.
unittest the spreadsheet-row to Sume bulk item builder, no network
A pure function that turns a spreadsheet row into a Sume bulk item, with four unittest cases for trimming, price format, a spend cap in range and blank SKUs.
Music Router image_url: public HTTPS only, and null to clear it
Sume music image_url must be a public HTTPS image, and null clears it on a reused request object. When a mood still helps and when the prompt does more.
Warranty registration how-to clip: avatar with your product image
A 20-second avatar clip can show owners how to register a warranty. Using the product image input, the per-second surcharge, script steps, and cost by tier.
Splitting a 30-second Wan 3.0 job adds cents from per-job rounding
Wan 3.0 bills linearly per second, but Sume rounds each job up to the cent. A 30 s clip as six 5 s jobs costs $3.78 at 720p instead of $3.75. Full split table.
Cancel a Format run: cancel_effect canceled vs no_op, and the bill
Cancel is idempotent. cancel_effect says canceled or no_op, a canceled run never sends a webhook, and generation that finished is still billed.
Cost of one finished 60-second ad: four Wan clips, voice, music
A finished 60-second ad built from four 15-second Wan 3.0 clips at 720p costs $7.9678 on Sume with voice, a music bed, one render and captions.
Retiring a webhook endpoint: Sume runs already carrying it still POST
Any run created with a webhook_url can POST when it ends, even after you decommission the endpoint. Retries run 10 times, and the receipt holds the real result.
Weekly toolbox talk video with an AI avatar: 52 clips a year, cost
A weekly 30-second avatar safety reminder costs a fixed sum per year at each tier. Script rules, review steps, and what it must not replace.
Keep a 2-minute video long-form on YouTube: render 1920x1080
Timeline 1.0 renders 1080x1920 by default. Set output width 1920 and height 1080 to keep a short video out of the Shorts class. YouTube's rule and the request.
Restore a scratched family photo in three staged edits
Fix creases, then scratches, then fading in separate Ideogram 4.5 passes. Three low-quality passes cost $0.1125 on Sume; a high final pass brings it to $0.35.
Founder story video with an AI avatar: who is on screen?
An About-us video built from an avatar raises one question: whose face. Three Sume avatar inputs for a founder story, a 40-second script, cost, and a label.
A job id is not a run id: poll image and video generation via /v1/jobs
Image and video generate routes create jobs, not runs. Poll GET /v1/jobs/:id/status until terminal is true. waitForRun is for run ids only.
Salon or barber intro clip: avatar in front of your own shop photo
Put an AI presenter in front of a photo of your salon with the scene photo input. Script, 9:16 default, what a 20-second clip costs, and a consent warning.
Node 22 script: create a Sume bulk queue and poll it to the end
Dependency-free Node 22 ESM script: POST a bulk queue from items.json, back off the poll, survive 429 and 503, and exit non-zero when any item failed.
Replace the presenter in a 3-minute training video with H3 Max Recast
Recast takes 5 to 30 seconds per job, so a 3-minute video is six chunks. Steps, the 768p cost, and why the original voice track becomes the timeline spine.
Outage status update video with an AI avatar: script and cost
A 20 to 30 second avatar clip can front a status-page update. What to script, what it costs by tier, and why the written update must go out first.
Midday to golden hour in a video: Omni edit or first/last frame
Change the time of day in a clip you already have. When a Gemini Omni Flash edit fits, when first and last frames fit, and what a 6-second pass costs on Sume.
Polling 200 video jobs every 30 s is 400 calls a minute: do this
A poll loop copied to Sume's 30 s example turns 200 open jobs into 400 status calls a minute. Use next_poll_after_seconds, backoff, or a webhook plus sweep.
Seedance 2.5 30-second 720p price by aspect ratio: token counts
On Sume 16:9, 9:16 and 1:1 all cost $17.34 for 30 s at 720p, while 21:9 costs $17.42 and 4:3 $17.45. Frame sizes, token counts and the formula.
Board-game crowdfunding trailer from box, minis and cards, Wan 3.0
Send up to 10 reference images of box, miniatures and cards to wan-3.0 for a 15-second trailer: $1.88 at 720p on Sume, with logo and card-text drift.
Microscope-style science clip with Gemini Omni Flash: prompt and cost
Google's own Omni 1.1 demo prompt is a diatom micrograph. Here is how to write one, draft it at 360p, and what 8 seconds costs on Sume at each resolution.
Add steam to a coffee or food clip with Omni edit, check the sound
Make a plain food or drink clip look hot by editing in steam with Gemini Omni Flash. Prompt wording, 8-second cost at 720p, and why the audio needs a listen.
Bash: split a TSV into 100-row Sume bulk queues with jq and curl
A short shell script that cuts a product TSV into 100-row chunks, builds each bulk body with jq, and posts it under a chunk-named idempotency key.
What Sume's 1.25 margin adds per 100 clips on Omni and Wan
Sume bills provider list times 1.25 on video-router models. 100 Omni 720p clips of 8 s carry $20 of margin; 100 Wan 1080p clips of 30 s carry $150.
Monthly investor update video with an AI avatar: 60 s per job
An avatar update longer than 60 seconds has to be split. How to cut a monthly investor or board update into three jobs, keep one presenter, and price it.
BigQuery product table to Sume bulk queues and a queue map
Read active SKUs from BigQuery, cast columns to plain JSON types, create one Sume bulk queue per 100 rows, and stream queue id, index and SKU to a map table.
Transcribe a file from your own bucket: what Sume STT audio_url needs
Sume speech-to-text takes a public HTTPS audio_url, preferably on media.sume.com. How to get a private recording ready, and what a neighbouring API rejects.
Remove a price sticker from a product photo and check the label
Remove a sticker from a product photo with Ideogram 4.5 on Sume, then check the label text underneath. A small mask on GPT Image 2.5 limits the edit.
Crop a 16:10 MacBook screen recording to 9:16: width 0.3516
A 16:10 recording needs crop width 0.3515625 for 9:16: 900x1600 at 2560x1600, but 674x1200 at 1920x1200 because 675 is odd. The request and the rounding.
Native 4K or 2K image edits: pixel ceilings by model
How many pixels each model reaches: Flux 3 Image near 16.8 MP, GPT Image 2.5 capped at 8,294,400 on Sume, Ideogram 4.5 at 2K. What Sume serves.
Sweep Format runs for failed webhook deliveries, then redeliver
Sweep in Python: read each run's webhook_delivery, redeliver only failed or exhausted ones, and leave the rest alone. Uses formats:read, formats:write.
Wan 3.0 bumper prices: 2 to 6 seconds at 480p, 720p, 1080p
A 2-second Wan 3.0 bumper costs $0.125 at 480p and $0.50 at 1080p on Sume. A 6-second one costs $0.375 to $1.50.
Format run cap headroom: usage.cap limit, counted and remaining
usage.cap on a run receipt splits the spend-cap check into limit, counted and remaining USD micros. Read it to see how close a run is to failing.
Route Sume run webhooks by event: format, action and agent terminal
Run webhooks use one terminal event per family: action.run.terminal, format.run.terminal, agent.run.terminal. Route on event, then branch on outcome.
Format run stuck on processing: stalled or slow? Read events_url
A Format run can show processing for a long time. The last at value on events_url is the progress clock: if it stops moving for minutes, the run is stalled.
Omni draft grid: four 360p variants, then one final. What it costs
Google's Draft Room idea, run through the Sume API: four 8-second 360p drafts that change one thing each, then a 1080p final. Total $2.70, with a script.
AI avatar vs a paid creator: break-even per 30-second clip
Divide a creator's quote by Sume's price for a 30-second avatar clip. Costs by tier for 1, 10 and 50 clips, with a retake allowance and what an avatar lacks.
Change a person's outfit in a video: Omni edit or H3 Max Recast
Two ways to change what someone wears in a clip on Sume: a text edit with Gemini Omni Flash, or Recast with a photo of the person. Rates for 10 seconds.
Audio detach size for a 180 s Short: 34.56 MB wav, 2.88 MB mp3
The byte arithmetic for a 3-minute track from audio detach: stereo 48 kHz wav, mono 16 kHz wav and 128 kbps mp3, with the fields that set each size.
Cut hold music from a call recording before STT: one concat job
Drop a hold-music stretch with one Sume timeline audio concat job that reuses the file twice via source_in and duration, then send the result to STT.
Sume media tools: which wait for a 200 and which always return 202
Video inspect defaults to sync; trim, filter, detach, compose and timeline to async; frames is always 202. The 30 s wait and where each result is read.