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
H3 Max 1080p regenerates from 768p: the price of the extra step
MiniMax says its 2K path regenerates in context. Sume documents H3 Max 1080p as a latent refinement of 768p, $0.20 vs $0.10 per second. When the doubling pays.
MiniMax H3 open weights: what is released and what stays hosted
MiniMax released H3-Base, FL2VA and Ref2VA checkpoints at 768p. Context-IR and the 2K regeneration module stay hosted. What that means for a Sume API user.
MiniMax says H3 is 2K: what Sume lists, 768p to 1080p
MiniMax's own page puts H3 at 2K and 15 s. Sume lists native 768p for H3 and H3 Max, 1080p refinement on Max, and priced 2K/4K upscale SKUs on H3. Prices.
Omni 1.1 Flash extends to 40 s; build 40 s on Sume from 4 clips
Google lists extension to 40 s for Omni 1.1 Flash. Sume's gemini-omni-flash-1.1 does 3-10 s per job, so 40 s is four jobs plus a Timeline join.
Is the Seedance 2.5 API live? What Sume can call for 30-second video
ByteDance's Seedance 2.5 post said its API was coming soon via BytePlus ModelArk. On Sume, seedance-2.5 (4-30 s) and wan-3.0 (2-30 s) are in the Video Router.
Omni Flash vs Omni 1.1 Flash: what changed and what Sume exposes
Original Omni Flash was 3-10 s at 720p; 1.1 adds 360p draft, 1080p, 4K, scene extension and frame control. What Sume's id exposes, read 2026-10-05.
Which Seedance for a UGC ad: 2.5, 2.0, Fast or Mini on Sume?
Test UGC hooks on Seedance 2.0 Mini or Fast at 480p, then spend on 2.5 only when the ad needs 16-30 s or 1080p. Prices for 15 s vertical clips.
'4k is not supported by Video Router v1': who takes 2K and 4K
Only minimax-h3 (2K and 4K upscales from native 768p) and gemini-omni-flash-1.1 (4K) accept those tiers on Video Router; every other model answers 400.
'720p is not supported by model minimax-h3; use 768p' and its twins
MiniMax H3 and H3 Max are 768p models on Sume: 720p gets a 400 that says use 768p. 768p is refused elsewhere, and 360p is Gemini Omni only.
AI photo edit came back a different size: four causes on Sume
A different edit size on Sume comes from four things: a shape request, a pixel string snapped to a ratio, the larger 1K edit size, or no aspect_ratio auto.
Filipino and Cantonese: fil and yue codes in MAI-Transcribe-2 vs Sume
MAI-Transcribe-2-Streaming lists fil and yue and treats unknown codes as unset. Sume takes a 2-16 character language_code hint; test a sample before relying.
Gemini Omni reference errors: 10 images, 3 clips of 3 s, 16:9 or 9:16
Gemini Omni Flash 1.1 on Sume refuses an 11th image, a 4th or longer-than-3-second clip, audio references, bitrate_mode and ratios other than 16:9 or 9:16.
generate_audio false on Gemini Omni: accepted, clip keeps its audio
Video Router accepts generate_audio false on gemini-omni-flash-1.1 and ignores it; MiniMax H3 answers 400. Check the model before you rely on a mute flag.
grok-imagine-video-1.5 on Sume: needs an image, seven fields refused
grok-imagine-video-1.5 is image-to-video only on Sume: send one image, no end frame, no reference video or audio, no aspect_ratio, no generate_audio.
Ideogram 4.5 21:9 returns 400 on Sume: use 2:1 at 2K and crop
21:9 is not an Ideogram 4.5 ratio on Sume and returns 400 with the supported list. Ask 2:1 at 2K (2880x1440), then crop to 2560x1080 and lose 225 px of height.
Ideogram 4.5 edit with resolution 2K and no ratio: still source size
On Sume, an Ideogram 4.5 edit with resolution 2K but no aspect_ratio keeps the source size. Add a ratio to get 2K pixels, such as 2560x1440 for 16:9.
Ideogram 4.5 pixel size to ratio: where 1920x1080 and 1600x500 snap
On Sume, an Ideogram 4.5 pixel size such as 1080x1350 or 2400x1000 is snapped to the nearest of 15 native ratios: 4:5 and 2:1. Table of seven common sizes.
Ideogram 4.5 output size per aspect ratio at 1K and 2K on Sume
Exact Ideogram 4.5 pixel sizes on Sume for all 15 ratios at 1K and 2K, plus the four ratios where a 1K edit is larger than text-to-image.
<IMAGE_REF_0> and <VIDEO_REF_0>: tag references in Gemini Omni
On gemini-omni-flash-1.1 you point at each reference by token in the prompt: <IMAGE_REF_0>, <VIDEO_REF_0>, numbered from 0 in list order. Request example.
Is Ideogram 4.5 edit a separate id on Sume? No, references decide
No. ideogram/ideogram-v4.5 is the only id: no input_references generates, 1 to 5 edits the first image. Size and aspect rules change with the call type.
'kling-3 supports text-to-video or start/end frames only': the fix
kling-3 on Sume has no reference lists. The 400 is fixed by sending text only, or a first frame and optional end frame, or by switching to a reference model.
MAI-Transcribe-2-Streaming has no server VAD: who ends the turn?
In the preview, turn_detection only accepts null, so your client sends the commit event. Sume STT has no live socket; it cuts sentences after the job finishes.
'reference_audio_urls requires a reference image or video'
Audio cannot be the only reference on Sume video models. Add one image or video reference. Which models take audio references, and their caps.
'Reference inputs must not exceed 12 total files': per-type caps
Most Sume video models take 9 reference images, 3 videos, 3 audios, 12 files total. Wan 3.0 takes 10, 5 and 5; Gemini Omni takes 10 images and 3 short videos.
Seedream 4.5 edit with n=4 returns one image on Sume
On Sume, a Seedream 4.5 call with reference images is an edit and always asks the provider for one image, so n=4 does not give four takes. Use four calls.
Video Router 'accepts a duration of at most N seconds': ceilings
The 400 names the model's own ceiling: 30 s on seedance-2.5 and wan-3.0, 15 s on most others, 10 s on gemini-omni-flash-1.1. Table of every floor and ceiling.
Gemini Omni video_url edit: the two 400s and how to avoid them
video_url on Video Router means edit, and only gemini-omni-flash-1.1 accepts it. Sending frames or references with it, or another model, gives a 400.
Wan 3.0 reference videos on Sume: 5 clips, 15 s total, 16 fps floor
Wan 3.0 takes up to 5 reference videos, 10 images and 5 audios on Sume; other models stop at 3 videos. Catalog notes: 15 s combined, at least 16 fps.
Wan 3.0 is a closed beta and API, not open weights: a Sume call
Alibaba's Wan 3.0 README points to wan.video and QwenCloud, not downloadable weights. Sume lists wan-3.0, 2 to 30 s. The request shape on Sume.
Gemini Omni Flash 1.1 on Sume: one model id, four request shapes
Sume routes gemini-omni-flash-1.1 by what you send: prompt, image_url, reference lists, or video_url. Fields, limits and the price of each in one table.
AI video resolution on Sume: 360p, 480p, 720p, 768p, 1080p or 4K
Which Sume video models take which resolution, and the per-second price at each tier. 768p is a native tier for MiniMax, not a 720p alias.
Is Seedance 2.5 open source? What the launch page says, and Sume
Searching Seedance 2.5 open source or Hugging Face? The launch page names Jimeng, Doubao Pro and a coming-soon API. How to call it on Sume.
AI video generator for YouTube: Sume models that do 16:9 at 1080p
Four Sume video models accept 16:9 at 1080p. Their per-second rates and the cost of a 5-second and 8-second B-roll clip at 1080p.
Text edits on Sume: use sume/auto or pin ideogram/ideogram-v4.5?
Sume's sume/auto never says which model ran, so a text-edit workflow that depends on Ideogram 4.5 behavior should pin ideogram/ideogram-v4.5.
Translate text in an image but keep brand names: the edit prompt
A prompt pattern for translating the words in an image on Sume while leaving brand names, prices and codes alone, with an Ideogram 4.5 request.
Edit the words on a neon sign photo with Ideogram 4.5 on Sume
Change the text on a neon sign in a photo with one Ideogram 4.5 edit call on Sume: the request, a prompt that keeps the glow, and what to check.
Omni Flash audio is always on, so there is no silent discount on Sume
Gemini Omni Flash 1.1 on Sume always renders native synced audio and rejects generate_audio false. The rate rows have no audio switch to save money.
Which AI video model for how many seconds: five length bands on Sume
Pick a Sume video model by clip length: 2 s, 3 to 4 s, 5 to 10 s, 11 to 15 s, and 16 to 30 s. Which ids accept each band, from the Video Router docs.
Ideogram 4.5 has two endpoints; on Sume it is one model id
Ideogram lists a precise edit endpoint and a generate-plus-edit endpoint. On Sume one id, ideogram/ideogram-v4.5, covers generate and edit by what you send.
Multi-turn edits or one combined prompt: Ideogram 4.5 restart rule
Ideogram says 4.5 removes artifact buildup so multi-turn editing works. When to chain edits on Sume, and when to restart from the original instead.
30-second ad in one request: which Sume video models accept 30 s
Seedance 2.5 and Wan 3.0 accept 30 seconds on Sume. MiniMax H3 stops at 15 and Gemini Omni at 10. A table of ranges, plus the vendor lengths that differ.
Omni Flash took 7 image references; Omni 1.1 on Sume takes up to 10
Google's original Omni Flash allowed up to 7 image references and 3 short clips. Sume's Omni Flash 1.1 route lists up to 10 images and 3 clips of 3 s.
How long can one AI video request be? Sume duration windows
Duration windows differ by model: 2 to 30 seconds on wan-3.0, 4 to 30 on seedance-2.5, 3 to 10 on Gemini Omni Flash 1.1. Read them from the catalog.
H3 Max Recast source rules: 5 to 30 seconds, no shot over 15
Recast swaps the people in a video for 1 to 4 photos. Source: 5 to 30 s, no shot over 15 s, duration set to the inspected length. $0.375 a second at 768p.
eleven_v4 vs eleven_v4_turbo: model IDs, endpoints, which to pick
ElevenLabs lists eleven_v4 for expressive speech with cloning in 90+ languages and eleven_v4_turbo at about 100 ms median latency. Which fits a video pipeline.
ElevenLabs languages: Flash v2.5 has 32, Multilingual v2 29, v4 90+
ElevenLabs lists 32 languages for Flash v2.5, 29 for Multilingual v2 and 90+ for v4 and v4 Turbo. Check your markets against the model, then log it per job.
Nano Banana reference image limits: Lite, 2 and Pro compared
Nano Banana 2 Lite takes up to 14 object images, Nano Banana 2 takes 10 object, 4 character and 3 style, Pro takes 6 object and 5 character.
Repeated image edits without drift: a native-resolution test plan
Recraft says Ideogram 4.5 is built for stable repeated edits at native resolution without downsampling. A five-pass test to check any edit model for drift.