Models
The models behind Sume's video, image, speech, and music endpoints: ids, clip lengths, resolutions, aspect ratios, audio, references, and migration.
Start with What AI models does Sume support? Every model id, by API surface
Does Lyria 3.5 audio carry a SynthID watermark?
Yes: Google's Gemini API music docs say Lyria audio carries a SynthID watermark. The Sume docs I read do not describe watermarking either way.
MAI-Transcribe-2-Streaming or a batch STT job: which fits?
MAI-Transcribe-2-Streaming returns first partials in just over 100 ms. Sume's speech-to-text is a batch job up to 10 minutes at $0.01 per minute.
Can I use MAI-Voice-2.1 audio commercially? What Microsoft says
Microsoft's MAI voices page says it holds full licensing rights for commercial use, the service is in public preview with no SLA, and cloning is gated.
21:9 ultrawide AI video API: which Sume models accept it
Seedance 2.5, Seedance 2 and its variants, MiniMax H3 and H3 Max accept aspect_ratio 21:9 on Sume. Wan 3.0, Kling 3 and Gemini Omni Flash do not.
Kling 4.0 Omni Reference takes 15 references; Sume limits compared
fal says Kling 4.0 Omni Reference accepts up to 15 references. On Sume, kling-3 takes none; Wan 3.0 and Gemini Omni Flash 1.1 publish their own limits.
Kling 4.0 Flash 3-second minimum vs Sume's 4-second kling-3 floor
Kling 4.0 Flash starts at 3 seconds. Sume's kling-3 starts at 4. The models that can render a 3-second clip on Sume, and what a 3 s request costs.
Swap the gift in a Christmas clip with Gemini Omni Flash edit
Gemini Omni Flash 1.1 on Sume edits a 3 to 10 second clip by prompt, so one approved Christmas ad can become three gift variants without a reshoot.
Which Korean voices does MAI-Voice-2.1 have, and with what styles?
MAI-Voice-2.1 lists four ko-KR voices: Grant, Harper (role styles), Haena (14 emotion styles) and Junho (12). Pair a voice with the right caption style.
MAI-Voice-2.1 English voices with call-centre or audiobook styles
customer_call_center and audiobook styles sit on en-US Grant, Harper, Sage and en-GB Emily, Harry. Iris, Jasper, Dhruv and Priya are neutral only.
MAI-Voice-2.1-Flash tops out at 45 seconds: what for longer?
MAI-Voice-2.1-Flash is reported at up to 45 s of audio and 150 ms latency. Longer lines need the standard model or a TTS taking 20,000 characters.
Runway Gen-4 Turbo takes an image only: text-to-video on Sume
Runway's Gen-4 Turbo accepts image input only. If your prompt has no source image, these Sume video models take text alone, and these two need a picture.
Runway lists Seedance 2.5 with video input: what Sume accepts
Runway says seedance2_5 takes text, image or video. On Sume, seedance-2.5 takes video as a reference, not an edit source. Which model edits a clip.
ElevenLabs Music v2.5 genre strengths and how to test them
ElevenLabs says v2.5 is strongest on vocal-led and acoustic-heavy genres such as R&B, soul, rock and orchestral. Test your genre with three briefs.
Choose an AI video model by the asset you already have: Sume ids
Photo, two frames, product shots, a voice track, a driving clip or a finished video: which Sume video id fits each starting asset, from the catalog rows.
Image reference limits: GPT Image 2.5 takes 16, Ideogram 4.5 takes 5
GPT Image 2.5 accepts up to 16 input references on Sume; Ideogram 4.5 edits the first image and uses up to 4 more. How to order them and when to prefer which.
MiniMax H3 references on Sume: 9 images, 3 videos, 3 audio, 12 total
minimax-h3 and minimax-h3-max accept 9 images, 3 videos and 3 audio files, 12 in total, and audio cannot be the only reference. Duration rules and errors.
Video aspect ratios by Sume row: Kling 3, Gemini, Wan 3.0, MiniMax H3
Which aspect_ratio values each Sume video row takes: kling-3 has three, Gemini two, Wan and MiniMax add adaptive, and MiniMax adds 21:9. From the catalog.
ChatGPT picks Flare or Sunburst for you; the API does not
In ChatGPT, Flare handles standard use and Sunburst takes complex prompts. In the API you choose. How to build that rule on Sume, where sume/auto uses Flare.
ElevenLabs Music stems: Creator 2 and 4, Pro 6, and the Sume route
ElevenLabs lists 2 and 4 stems on Creator and up to 6 on Pro. When a video editor needs stems, and when a Sume duck_db setting does the job.
Omni Flash 1.1: GA in the Gemini API, Preview on Agent Platform
Gemini API release notes record gemini-omni-1.1-flash as GA on Aug 27, 2026; an Agent Platform page title still says Preview. What to check first.
Omni's 1M-token context: how many ten-second edit turns fit?
Omni lists a 1,048,576-token context and 5,792 tokens per video second. If prior clips stay in context, 18 ten-second turns fit; Sume edits are stateless.
Google Lyria ids and prices next to the Sume Music Router ids
Google lists Lyria 3.5 at $0.08 a song, a 30-second clip model at $0.04, and Lyria 3 Pro Preview at $0.08. Which of them Sume routes, and at what flat price.
Sume motion routes by length: Kling, Recast and Genjutsu
Three Sume routes drive a result from a source clip, each with a length rule: Kling takes duration_seconds, Recast follows the source, Genjutsu a range.
Suno v6 licence deals: Warner, BMG, Believe, with dates to log
Suno v6 was announced Sept 9, 2026 after deals with Warner (Nov 25, 2025) and BMG (Aug 12, 2026), with Believe named as a partner. What to log for an ad file.
Wan 3.0 reference videos: 15 seconds total and a 16 fps floor on Sume
Sume's wan-3.0 row takes 10 images, 5 videos and 5 audio files, but videos and audio each cap at 15 seconds combined, and video needs 16 fps.
YouTube Shorts' free Veo 3 Fast clip is 480p: what Sume models render
YouTube's 2025 announcement put Veo 3 Fast in Shorts at 480p with sound. Sume's router lists 720p and 1080p models; check the ceilings before you pick.
YouTube Speech to Song vs Sume Music Router: prompt in, audio out
YouTube's Speech to Song turns a Short's speech into a song with Lyria 2. Sume's Music Router takes a text prompt (and an optional image), with no audio input.
1-second AI video API: nothing on Sume goes below 2 seconds
No Sume video model takes duration 1. The floor is 2 seconds on Wan 3.0; the rest start at 3, 4 or 5. How to get a 1-second beat anyway, with trim and cost.
2-second AI video API: only Wan 3.0 takes it on Sume
A 2-second clip on Sume means Wan 3.0: $0.13 at 480p, $0.25 at 720p, $0.50 at 1080p per job. Which modes work at 2 seconds and where the other models start.
31-second AI video API: why Sume needs two jobs past 30 seconds
No Sume video model takes duration 31. Wan 3.0 and Seedance 2.5 stop at 30 seconds; here is the two-job route, what it costs and how to join the clips.
Kling 4.0 Flash is first-frame only; what Sume's Kling 3 accepts
Kling 4.0 Flash early access takes a first-frame image, 20 s, 720p, SDR. Sume ships kling-3: 4 to 15 s, 720p/1080p, priced per second. Limits side by side.
Seedance 2.0, Fast and Mini next to 2.5: still callable on Sume
Sume still serves seedance-2, seedance-2-fast and seedance-2-mini by id next to seedance-2.5. 15-second ceiling, per-token prices, 4 s and 5 s job costs.
How long can a Seedance 2.5 product clip be on Sume?
On Sume, seedance-2.5 accepts 4 to 30 seconds at 480p, 720p or 1080p. How to pick a length for a product clip and poll the job until it completes.
Check Sume image-model ratios in code before a channel render
Before rendering one product photo for four marketplaces, read GET /v1/images/models and pick a model whose aspect_ratio values cover every channel you sell on.
Index-Translate 2B, 9B or 35B-A3B for subtitle translation
Index-Translate ships 2B, 9B and 35B-A3B text models. What the vendor pages say about each size, and how to pick one for subtitle cues you burn with Sume.
Index-Translate, Echo, Homura: which part do you need?
Bilibili's Index-Translate release has text, speech, dubbing and length parts. Which fit subtitles, and which Sume step follows.
Gemini Omni has no 1:1: cut the square Google Ads wants with Sume
Asset Studio exports 16:9, 9:16 and square, but Gemini Omni on Sume renders only 16:9 or 9:16. Crop the 1:1 with video-filter: x 0, y 0.21875, 1 by 0.5625.
Ideogram 4.5 is coming to Recraft Studio: the API route for edits
Recraft says Ideogram 4.5 is coming to Recraft Studio for drift-free edits. If you need it in code today, Sume lists ideogram/ideogram-v4.5 with edit support.
Kling 3.0 element references are not on Sume's kling-3: use these
Kling's guide lists start frame plus element reference for Kling 3.0. Sume's kling-3 refuses reference inputs; Seedance 2.0 and Wan 3.0 accept them.
Veo 3.1 Lite takes text and image only; Sume models take more inputs
Google lists Veo 3.1 Lite inputs as text and image. On Sume, Seedance 2.0 and Wan 3.0 also accept video and audio references; Kling 3 does not.
Do you need streaming transcription? A recorded clip needs a batch job
MAI-Transcribe-2-Streaming is built for live audio. For a finished clip, Sume STT at $0.01 a minute returns text and word timings. How to decide which you need.
Sume Auto or a pinned model for repeatable image edits?
With sume/auto the response echoes sume/auto and never names the family that ran. For edits you must reproduce, pin a catalog id such as openai/gpt-image-2.5.
Which Kling 3.0, Seedance 2.0 and Wan 3.0 ratios count as a Short?
YouTube makes a Short from a square or vertical video up to 3 minutes. Of Sume's ratios, 9:16, 3:4 and 1:1 qualify; 16:9, 4:3 and 21:9 do not.
Mercury Voice vs text to speech: what a voice agent LLM does not do
Inception's Mercury Voice writes replies for voice agents. It does not synthesize audio. Where TTS and STT jobs sit around it, with Sume rates for each.
Draw boxes on a photo, then edit it: Sketch-style markup on Sume
ChatGPT Sketch is a drawing panel inside ChatGPT. On the API you get the same effect by sending a marked-up photo as a reference and saying what the marks mean.
FLUX 3 docs show 67.8 to 94.1% unchanged pixels: measure your edits
BFL's FLUX 3 Image docs publish pixel-difference examples, not a guarantee. Measure the unchanged-pixel share of your own Sume edits with a 20-line script.
Ideogram 4.0 claims 0.97 OCR accuracy: measure your own text score
Ideogram's 0.97 English OCR figure is a vendor benchmark. Sume serves Ideogram 4.5, not 4.0, so run 20 of your own headlines through OCR and count matches.
Kling 3.0 allows 3 seconds at Kling; Sume's kling-3 starts at 4
Kling's guide lists 3 to 15 seconds for Kling 3.0. Sume's kling-3 takes 4 to 15, so a 3-second beat means a 4-second clip, then a trim.