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
AssemblyAI Universal-3.6 Pro 32 languages vs Sume STT language_code
AssemblyAI streaming defaults to Universal-3.6 Pro with 32 languages; 3.5 Pro has 19. Sume STT 1.0 takes one optional language_code hint and auto-detects.
Sonic 3.6 written disfluencies: pacing and Sume's emotion guide
Sonic 3.6 shifts pacing for written hesitations like uh. How to write a natural-sounding transcript, and what Sume's optional emotion guide adds.
Cartesia STT keyterms for brand names, and Sume STT without keyterms
Cartesia STT adds keyterm prompting for brand names and invented words. Sume STT has no keyterm field, so here is how to fix brand names after transcription.
Deepgram Nova-3 de-CH, Lithuanian, Portuguese: language_code on Sume
Deepgram improved Nova-3 for Swiss German, Lithuanian and Portuguese on September 24. What to send as language_code on Sume STT for regional variants.
FLUX 3 Action is a robot-control model, not an image model
FLUX 3 Action is an open-weight 7B world action model for robots. It predicts actions, not images, so it is no drop-in for the FLUX ids Sume serves.
FLUX 3 Image bounding boxes vs Sume masked edit: what each takes
FLUX 3 Image takes a JSON list of boxes at the end of the prompt. Sume's image edit takes a mask image: mask_image_url with image_urls. Differences, no overlap.
FLUX 3 Image 10 reference images vs Sume's 1-10 image_urls
FLUX 3 Image takes up to ten references named by prompt position. Sume's image edit takes one to ten public image_urls, with no FLUX 3 id in the docs.
FLUX 3 x mimic: one backbone for video, audio and robot actions
BFL says FLUX 3 is one model trained jointly on images, video and audio, and FLUX-mimic adds robot actions. What that does and does not tell a media API user.
FLUX 3 video continuation caps at 15 s: extending a clip on Sume
BFL cut FLUX 3 video continuation (v2v) to 15 seconds on Aug 17; other modes stay at 20. Sume has no continuation mode: chain a last frame instead.
FLUX 3 Video 4K uhd (3840x2176) and which Sume video models do 4K
FLUX 3 Video's uhd resolution is 3840 x 2176 at 16:9. On Sume, a 4K request goes to a model whose supported_resolutions lists it; here is what the docs say.
Gemini 3.8 Live Extended Thinking: the id, and Sume's audio ids
gemini-3.8-live-extended-thinking is a Live API audio-to-audio model id. Sume has no live audio route; its audio ids are sume/music-auto and TTS Router ids.
Grok Imagine 21:9 and 5:2 ratios vs Sume's Grok aspect list
xAI added 21:9 and 5:2 to Grok Imagine image on Aug 28. Sume's Grok list has 20:9 and 2:1 but neither of those two, so pick a listed ratio for banners.
Genjutsu Restyle's 20+ styles vs Sume motion transfer inputs
Higgsfield Genjutsu Restyle picks from 20+ styles in its app. Sume's motion transfer takes a source video plus 1 to 8 reference images, with no style list.
Kling 4.0 on-screen text vs Sume burned-in captions
Kling 4.0 renders on-screen text inside the video. Sume burns captions onto a finished clip; language is a speech-to-text hint, and cues add authored text.
Kling 4.0 video extension, forward or backward, done by hand on Sume
Kling 4.0 extends a video forward or backward in one workflow. On Sume you chain clips yourself: pull a frame, pass it as first_frame or last_frame.
Kling Avatar output length follows audio; Sume H3 Max is 5-14.8 s
On fal, Kling Avatar v2 Pro output duration matches the audio file. On Sume the audio sets length too, within route windows from 5 to 14.8 s up to 60 s.
Kling Avatar v2 Standard vs Pro: fal tiers vs Sume quality
fal prices Kling Avatar v2 Pro at $0.115/s and Standard at $0.0562/s. Sume Avatar Video has its own quality values: standard, plus and max.
MiniMax H3 Max reference audio needs an image or video on Sume
Krea says MiniMax H3 Max audio references cannot be sent alone; Sume's reference_audio_urls has the same rule. Limits for images and videos compared.
Wan 3.0 references or start/end frames, not both: two Sume calls
Krea says a Wan 3.0 job takes reference media or start/end frames, not both. Sume lists them as separate modes, so split the work into two requests.
LTX-2.5 Cinemagraph LoRA vs image-to-video on Sume
The LTX-2.5 Cinemagraph LoRA is an image-to-video adapter for selective motion. On Sume, start from first_frame and describe the motion in the prompt.
Luma layers API: 10 RGBA layers vs Sume's flat image result
Luma type layering splits one image into up to 10 ordered RGBA PNG layers. Sume returns flat images at data[].url; n counts images, not layers.
Luma Scenes Ray 3.2 or Seedance 2: pin a model or use auto
Luma says Ray 3.2 and Seedance 2 are not interchangeable. On Sume, pin a catalog model for a known clip length, or send sume/auto and let Sume pick.
Luma uni-1 API: image models, types, and what Sume lists
Luma's API has two image models, uni-1 and uni-1-max, each with image and image_edit types. Sume lists its own catalog; check the models endpoint.
MAI-Transcribe-2 languages (60) vs Sume STT language hint
Microsoft's MAI-Transcribe-2 preview covers 60 languages. Sume STT 1.0 takes one optional language_code hint and auto-detects when you omit it.
MAI-Transcribe-2 speaker diarization vs Sume STT: no speaker field
MAI-Transcribe-2 attributes each segment to a distinct speaker. Sume STT has no diarize field: provider knobs are fixed server-side and you get words[].
MiniMax-H3 Fun ControlNet Union 2.0: 8 conditions vs Sume references
MiniMax-H3-Fun-Controlnet-Union-2.0 adds Scribble, Layout and Gray to five older conditions. Sume's minimax-h3 ids take image, video and audio references.
MiniMax H3 license for an EU company: contact MiniMax first
MiniMax H3 license Section II invites people in the EU, UK, Korea and USA to contact MiniMax about a license. What it promises and omits.
MiniMax H3 license: EU, UK, Korea and US are Excluded Territories
The MiniMax H3 Community License defines the EU, UK, Republic of Korea and USA as Excluded Territories and says use there is not authorized. What it quotes.
Muse Voice Transcribe 25+ languages and code-switching vs Sume STT
Muse Voice Transcribe supports 25+ languages and code-switching in one conversation. For mixed-language audio on Sume STT, omit language_code or set one hint.
Muse Voice Transcribe keyword biasing vs Sume STT: no vocab field
Muse Voice Transcribe has keyword biasing for names and domain terms. Sume STT has no vocabulary field: its request takes a language hint and word timings.
Pika Camera Director vs a prompted video edit on Sume
Pika Camera Director reshoots a scene from new angles. On Sume, Omni Flash 1.1 edits a clip from a video_url and a prompt; there is no camera-angle parameter.
Invisible models in Pika vs Sume Auto echoing sume/auto
Pika hides model names unless you ask. Sume's model sume/auto also hides the family and echoes sume/auto; pin a catalog id when you need a known model.
Pika Soundtrack: video in, sound out. What Sume has instead
Pika Soundtrack scores a finished video. Sume's docs offer generate_audio at video creation and a Timeline soundtrack bed, but no video-to-sound endpoint.
30-second video with sound: Pika Video Studio vs Sume ids
Pika Video Studio makes up to 30 seconds with sound. On Sume, seedance-2.5 and wan-3.0 reach 30 seconds; every other catalog model stops at 15.
Pika Video Studio's 50 references vs Sume's per-model limits
Pika Video Studio attaches up to 50 references. Sume's caps are per model: 10 images and 3 short clips on Omni Flash, 1 to 8 images on Genjutsu.
Qwen-Image 2.1 ControlNet Union Fun in ComfyUI vs reference images
ComfyUI added Qwen-Image 2.1 ControlNet Union Fun for editing on 2026-09-29. Sume has no ControlNet input; it edits with reference images and masks.
Qwen Image 2.1 license: research only, commercial needs a request
Qwen-Image-2.1 ships under the Qwen Research License: non-commercial means research or evaluation only, and commercial use needs a separate license.
Runway Compositing Nodes vs Sume Timeline compose
Runway added compositing nodes to blend images and video in Workflows. Sume's Timeline compose puts one still and one video in one MP4: stack or overlay.
Synthesia Express-3 follows script sentiment; Sume's script controls
Synthesia Express-3 avatars follow script sentiment. Sume Avatar Video is steered by script text, silence beats, scene direction and quality tiers.
Tavus 50 new Phoenix-4.5 stock faces vs a Sume avatar handle
Tavus added 50 first-party Phoenix-4.5 stock faces, 28 of them Pro. On Sume you create your own avatar from a prompt, profile or image and reuse its handle.
VEED lipsync-v2 video + audio vs Sume still + audio lip sync
fal lists veed/lipsync-v2 as video plus audio in. Sume lip sync routes start from a still and an audio clip, so existing footage is not re-synced.
Wan-Animate-2 14B: Apache 2.0 weights vs Sume Motion Transfer
Wan-Animate-2 14B weights are Apache 2.0 and ship inference scripts. Sume's Motion Transfer takes one source video plus 1-8 reference images, 4-30 s.
Wan-Animate-2 Base vs Distillation weights: which to download
Wan-Animate-2 ships Base and Distillation checkpoints, each with its own YAML config. What the card says, and the hosted limits Sume lists for comparison.
WorldCrafter camera control: Base weights, adapter and LoRA
TencentARC WorldCrafter-Base ships transformer weights with a matching camera adapter and LoRA. On Sume, steer camera motion in the video prompt instead.
YuE2-3B license is CC-BY-NC-4.0: what that means vs Music Router
YuE2-3B is tagged cc-by-nc-4.0, a non-commercial license. Sume's Music Router routes sume/music-auto, lyria-3.5 and lyria-3-pro instead.
YuE2 48 kHz stereo on a 24GB GPU vs a hosted Sume music job
YuE2-3B needs a 24GB NVIDIA GPU with BF16 for 48 kHz stereo songs. A hosted Sume music job needs no GPU, only a prompt and an Idempotency-Key for retries.
Suno v6-wild: is there a less predictable setting on Sume?
Suno v6-wild is a paid model that is less predictable. Sume's music request has no seed, temperature or guidance field, so variety comes from the prompt.
Lyria 3.5 takes 10 images: Sume music takes one image_url
Google's Lyria 3.5 page lists up to 10 images alongside text. Sume's music request has a single optional image_url, so pick one accepted still per track.