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
Gemini Omni rapid-fire video: a new labelled item every second
Prompt Gemini Omni for a rapid-fire clip that shows a different item every second with a text label, then send it through Sume's video router in 9:16.
Vietnamese, Thai, Indonesian, Malay TTS API: vi, th, id, ms on Sume
Cartesia Sonic 3.6 lists vi, th, id and ms. How to send each as the language on Sume TTS, why the field is required, and a per-character cost for each script.
GPT-5.1, o3, GPT-5.4 Nano shutdown dates and the Sume model field
OpenAI lists shutdowns for gpt-5.1, gpt-5.3-codex, gpt-5.4-nano (Apr 1, 2027) and o3 (Dec 11, 2026). A Sume Format run with those ids gets a 400.
German and Italian text to speech API: de and it on Sume TTS
German (de) and Italian (it) are in both Cartesia Sonic 3.6 and Sume's voice library. Send language de or it, reuse one voice, and watch the 409 language check.
ACE-Step 1.5 vocal-to-background music: what Sume does instead
ACE-Step 1.5 lists a vocal-to-background-music mode, covers and repainting under MIT. Sume's music jobs are prompt-only, so here is how to get a vocal-free bed.
AI video models with 4:3 and 3:4 aspect ratios on Sume
Seedance 2.5 and 2.0, Wan 3.0, MiniMax H3 and H3 Max list 4:3 and 3:4 on Sume; Kling 3.0 and Gemini Omni Flash 1.1 do not. Prices for the per-second rows.
Which aspect ratios does each Sume image model accept? Full matrix
A model-by-ratio table from the Sume image catalog: 1:1 to 21:9 plus 3:1, 4:1 and 8:1. Which rows accept 2:3, 21:9, 3:1 or auto, and which return 400.
How many reference images does each Sume image model take?
Sume's image catalog lists 16 references for ChatGPT Image 2.5, 5 for Ideogram 4.5, 10 for most others and 0 for five text-only models. Full table.
Kling 3.0 square video API: can I request 1:1 on Sume?
Yes: Sume's kling-3 row lists 16:9, 9:16 and 1:1 at 720p or 1080p, 4 to 15 s, with an audio-on rate. What it refuses, what it costs, and a request body.
Seedream 4.0 image API on Sume: price, ratios, 10 references
Seedream 4.0 (bytedance-seed/seedream-4) lists $0.0325 per image on Sume, accepts auto aspect ratio and up to 10 reference images. Request and limits.
Shortest AI video clip each Sume model accepts, and its price
Wan 3.0 starts at 2 seconds ($0.125 at 480p), Omni Flash at 3, Seedance 2.5 and Genjutsu at 4, H3 at 5. Minimum request per model, priced.
Which AI video models accept an input video on Sume?
Seedance, Wan 3.0, H3, H3 Max and Gemini Omni Flash take video references; Recast, Genjutsu and Omni edit need a source video. Kling and Grok take none.
Which AI video models take a reference audio file on Sume?
Seedance 2.x, Wan 3.0, MiniMax H3 and H3 Max accept a reference audio clip on Sume; Gemini Omni Flash, Kling, Grok, Recast and Genjutsu do not.
FLUX.2 pro has seed control: why a seed fails on Sume
fal lists seed control for FLUX.2 pro. Sume's docs say seed is schema-only today and returns 400, so here is how to get close to a repeatable result.
Google Pics API? Edit one object with Nano Banana on Sume
Google's Pics announcement describes an app, not an API. The closest call on Sume: Nano Banana reference edits, or a GPT Image 2.5 mask for one region.
Grok Imagine Image 2.0 returns up to 10 images: n on Sume
xAI's docs say up to 10 images per request at a flat per-image price. On Sume, the Grok row has its own n ceiling, so read it before you batch.
Read the Seedance 2.0 catalog row on Sume, field by field
The sample row from GET /v1/videos/models explained: durations, ratios, frame and reference types, audio, seed, pricing_skus and the empty passthrough list.
Recraft V4.1 Utility vs standard: which line for product mockups
Recraft's docs split V4.1 into standard and Utility lines, both $0.035. When flat, front-facing output helps, and what Sume's catalog lists today.
Seedream 5.0 Lite makes 1-6 images per run: how n works on Sume
fal lets Seedream 5.0 Lite return 1 to 6 images per generation. On Sume, n is capped per model in the catalog, so read the range before you batch.
Seedream 5.0 Lite 9MP (3072x3072) vs Sume's resolution tiers
fal lists Seedream 5.0 Lite up to 9 megapixels at $0.035. Sume takes tiers (512, 1K, 2K, 4K), so read the model's resolution descriptor before pinning one.
16:9 avatar video for slides and webinars: aspect ratio on Sume
Sume avatar video defaults to 9:16 but supports 16:9 and 4:3. How to set it for slides and webinars, with a 45-second Plus cost of $11.03 and caveats.
21:9 architecture photos with AI: which Sume models take the ratio
Nano Banana 2 and Pro, FLUX.2 pro and flex, Qwen and Recraft list 21:9 on Sume. GPT Image 2.5 gets 3840x1648 by custom size. Imagen, Ideogram, Seedream do not.
Edit a video with a prompt and a photo: Omni edit takes no references
Gemini Omni Flash 1.1's video edit on Sume takes a video_url and a prompt only. To put a specific person from a photo into a clip, use h3-max-recast instead.
How many Seedance 2.5 clips fill a 3-minute Reel: six at 30 s
Seedance 2.5 on Sume takes 4-30 s per clip, so a 3-minute Reel needs at least six. The clip math, the stitch step and Instagram's 3-minute guidance.
Imagen 4 has no 4:5 ratio: get Instagram portrait from 3:4 on Sume
Imagen 4 Fast and Ultra on Sume list five ratios and none is 4:5. Generate 3:4 and trim 90 pixels, or pick a model that lists 4:5. A Pillow crop included.
MAI-Voice-2.1-Flash makes 45 s of audio: narrating a 5-minute script
Microsoft lists 45 seconds of audio per Flash generation. A 5-minute script needs about seven chunks. A Python splitter, and how Sume handles long text.
MAI-Voice-2.1 on OpenRouter and Vercel: is it in Sume's TTS Router?
Microsoft lists MAI voices on Foundry, OpenRouter and Vercel. Sume's TTS Router lists only Cartesia Sonic ids today. How to check the live catalog.
Music Router routed_model: log which engine made each track
With sume/music-auto, job.request.routed_model names the engine that ran, such as lyria-3.5. Why to store it per track and how to read it from a job.
Seedance 2.5 video-to-video: a reference clip is not a person swap
Seedance 2.5 on Sume takes reference videos to condition a new 4 to 30 second generation. To swap people in an existing clip, send it to h3-max-recast instead.
Video-to-video clip length limits on Sume, by tool
Recast takes 5 to 30 s, Genjutsu 4 to 30 s, Omni edit 3 to 10 s, Avatar face swap about 4 to 15 s. A table and a Python picker choose by clip length.
What happens to the audio in each Sume video-to-video tool
Recast and Avatar Face Swap keep the source audio, Kling motion control keeps it by default, Omni always produces native audio, Genjutsu takes no audio field.
Where to try MAI-Voice-2.1 before paying: a 10-minute listening test
Microsoft lists the MAI Playground, Copilot Audio Expressions and Foundry for MAI-Voice-2.1. Run a ten-minute listening test with a fixed script and cost it.
FLUX.2 pro prompt upsampling is on by default; Sume has no switch
fal's FLUX.2 pro page says automatic prompt upsampling is on by default. Sume exposes no passthrough field for it, so test exact-text prompts with n=4.
4:5 on FLUX.2 pro and Qwen Image becomes 1229x1536 on Sume
On Sume, FLUX.2 pro, Qwen Image and Recraft V4 turn custom ratios like 4:5 and 21:9 into pixels with a 1536 long side. The sizes table, and a 1080x1350 resize.
Gemini Omni Flash camera words: smartphone zoom, film, whip pan
The camera and shot vocabulary Google's Omni guide lists, from locked off to dolly zoom and webcam style, with a Sume test grid to compare them cheaply at 360p.
Gemini Omni Flash on-screen text: the four things to specify
Google's Omni guide says to give text type, placement, animation and exposure. Prompt template, a Sume request, and why to check spelling before publishing.
Gemini Omni Flash adds cuts: how to prompt a single shot
Omni Flash tries a few shots by default. Google's wording for one unbroken take, and the same request on Sume's Video Router with a check for hidden cuts.
Higgsfield Soul 2 custom_reference_id is not a Sume parameter
Higgsfield's Soul 2 API takes custom_reference_id for a saved character. Sume's Soul is text-only; for a consistent character use a model that takes references.
Soul 2 lookbook batch of 4: Higgsfield Soul on Sume
Generate a four-image lookbook set per request with Higgsfield Soul 2: batch_size 1 or 4, 720p or 1080p, seven ratios, text-to-image only.
Kling 3.0 in the Sume Videos panel: audio on or off cost per clip
The Videos panel's Create estimate prices Kling 3.0 at $0.14 per second with audio off and $0.21 with audio on: 5 s is $0.70 or $1.05; 10 s is $1.40 or $2.10.
21:9 on kling-3 returns 400: which Sume video ids accept ultrawide
Kling's page lists 21:9 output for Kling 4.0. On Sume, kling-3 accepts only 16:9, 9:16 and 1:1; seedance-2.5 and the MiniMax H3 ids accept 21:9.
Kling 4.0 takes 15 reference assets; what Sume models accept
Kling's page lists up to 15 combined references for Kling 4.0. On Sume the cap is per model: Wan 10/5/5, H3 9/3/3 (12 total), Omni 10 images and 3 short videos.
Kling 3.0 starts at 3 seconds; Sume's kling-3 starts at 4
Kling's own page lists 3-15 seconds for Kling 3.0. The kling-3 id on Sume accepts 4-15, so a 3 second request returns unsupported_capability.
Luma Ray3.2 makes 5 or 10 second clips; which Sume ids accept both
Luma's page lists 5 or 10 seconds native for text and image to video. On Sume, seedance-2.5, wan-3.0, kling-3, minimax-h3 and Omni all accept both 5 and 10.
MiniMax H3 2K and 4K upscale on Sume: H3 accepts them, H3 Max does not
minimax-h3 will price a 2K or 4K request even though its resolution list shows only 480p and 768p. minimax-h3-max rejects both. Costs for 5 to 15 seconds.
Qwen Image Max is text-only on Sume; Qwen Image takes references
Qwen Image Max on Sume rejects reference images. For edits use qwen/qwen-image. The two ids compared: references, n, ratios, and the $0.075 Max list price.
Recraft V4 packaging concept art: text-only, WebP on Sume
Use Recraft V4 to explore packaging looks before a shoot: vendor price and design claims, plus Sume's text-only, WebP-only catalog entry.
Seedance 2.5 'extend twice': what Sume's 30-second request does
ByteDance says Seedance 2.5 makes up to 30 seconds with two extensions. Sume's seedance-2.5 id takes 4-30 seconds per request and has no extend field.