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
Mandarin Chinese text to speech API: MAI-Voice-2.1 zh-CN vs Sume zh
MAI-Voice-2.1 lists Chinese (Simplified) as zh-CN; Sume tags zh and bills by character. Audition Mandarin ad copy for 1 cent and know what to check.
Spain vs Mexico, Brazil vs Portugal: regional voiceover on MAI, Sume
MAI-Voice-2.1 lists es-ES, es-MX, pt-BR and pt-PT. Sume's guard compares only the primary language. How to pick a regional read and test it for 1 cent.
Tavus Human Interaction Model: what Griffin is, and what ships today
Tavus calls Griffin a Human Interaction Model. Only select testers have Griffin-Lite. Here is what it is, and what ships today on Sume for a talking presenter.
What is a full-duplex AI avatar, and when is a rendered clip enough?
A full-duplex avatar listens and talks at once, as in Tavus Griffin-Lite. If your viewers never talk back, a rendered Sume avatar clip does the job.
Thai, Vietnamese, Indonesian voiceover: on MAI's list, not Sume's tags
MAI-Voice-2.1 lists th-TH, vi-VN and id-ID; Sume's 16 voice tags do not. What that means, a 1-cent audition, and the Unicode trap in Vietnamese.
Czech, Hungarian, Romanian, Polish TTS: what MAI lists, what Sume tags
MAI-Voice-2.1 lists cs, hu, ro and pl. Sume tags only pl. A one-cent test plan and the Romanian s-comma versus s-cedilla encoding trap that breaks reads.
MAI-Voice-2.1-Flash at 45 ms: does a rendered avatar need fast TTS?
Microsoft lists MAI-Voice-2.1-Flash at about 45 ms of inference. A rendered avatar clip does not benefit from it. Where the latency shows up in a Sume job.
Seedance 2.5 21:9 is 1470x630 at 720p: size, price and request
Kling 4.0 adds 21:9. On Sume, Seedance 2.5 already makes 21:9 clips: 1470x630 at 720p, $5.81 for 10 s. Pixel sizes, prices and the request.
kling-3 has no reference_video_urls; motion_video_url is another route
Sume's kling-3 id rejects reference_*_urls. A clip that should drive a character's motion goes to the separate Kling motion control route as motion_video_url.
Gemini Omni video in the Windows app, and by API on Sume
Google lists Gemini Omni video in the Gemini app for Windows. For your own app, Sume exposes gemini-omni-flash-1.1 at 360p to 4K, 3 to 10 seconds.
Choose a Sume video model in five questions
Five questions pick a Sume video model: length, resolution, source media, audio toggle and price. One dated table maps each answer to model ids.
Kling 4.0 Auto aspect ratio vs Sume's explicit aspect_ratio field
Kling 4.0 lists an Auto format next to 16:9, 1:1, 9:16 and 21:9. Sume's video ids take explicit ratios; see which ids list which values and how to read them.
MiniMax H3 Max: five free fal runs a day, then the Sume cost per clip
fal's page offers five free H3 Max generations a day of up to 15 seconds. Use them to test a prompt, then price the paid clip on Sume at 480p, 768p or 1080p.
higgsfield-genjutsu missing from /v1/videos/models: why
higgsfield-genjutsu is in the Sume video catalog only when its provider is configured. It is Motion Transfer: one video_url plus 1-8 images, 480p or 720p.
/v1/videos audio_url references: Seedance 2 yes, Omni and Kling no
Only models whose supported_input_references lists audio_url take audio references on Sume: Seedance 2.x, Wan 3.0 and MiniMax H3 yes; Omni and Kling no.
Gemini Omni Flash edits 10-second inputs; Sume uses video_url to edit
Google's Gemini Omni Flash edits video inputs up to 10 seconds. Sume routes gemini-omni-flash-1.1 edits through a video_url, with no duration or aspect_ratio.
GPT Image 2.5 can take 2 minutes: submit async, not sync, on Sume
OpenAI says complex GPT Image 2.5 prompts can run up to 2 minutes. Sume's image route blocks only 30 seconds, so send mode async and poll the job.
GPT Image 2.5 edits through Sume: mask_url and 16 references
Sume's GPT Image 2.5 takes up to 16 references, an optional mask_url and background auto, transparent or opaque. OpenAI: transparent needs png or webp.
Wan 3.0 inputs: text, image, video, audio on Alibaba vs Sume wan-3.0
Alibaba documents text, image, video and audio inputs for Wan 3.0, up to 30 s. Sume's wan-3.0 does text, image plus end frame, and references, 2 to 30 s.
Every Lyria 3.5 track carries SynthID: what brands should know
Google says all Lyria 3.5 output carries a SynthID audio watermark and blocks artist voices and copyrighted lyrics. Here is what that means for brand music.
FLUX 3 Image open weights are 'coming weeks': API now or wait?
Press says FLUX 3 Image open weights arrive in the coming weeks; BFL lists commercial weights by licence. Call an API now if you need results this month.
Nano Banana 2 Lite is 1K only: send 2K and 4K to NB2
Google lists Nano Banana 2 Lite at 1K only; NB2 offers 512px to 4K at $0.045 to $0.151 per image. On Sume, resolution accepts 512, 1K, 2K or 4K per model.
Nano Banana 2 Lite has no Search grounding; NB2 does
Google's docs say Nano Banana 2 supports Google Search and image search grounding; Lite does not. Sume has no grounding flag and no Lite row.
Image launches this week vs Sume's catalog: FLUX 3, Ideogram 4.5
GPT Image 2.5 (Sep 22), Seedream 5.0 Flash (Sep 23), Ideogram 4.5 (Sep 30), FLUX 3 Image (Oct 1): Sume lists GPT 2.5, Ideogram 4.5 and Seedream 5 Lite.
Nano Banana 2 Lite takes 14 references; Sume's cap is 10
Google's docs give Nano Banana 2 Lite 14 reference images in total. Sume's input_references cap is 10 (16 on ChatGPT Image 2.5), and Lite isn't in its catalog.
Qwen-Image 2.1 native RGBA vs Sume's ChatGPT Image 2.5 transparency
Qwen-Image 2.1's card lists native RGBA transparency under a research licence. On Sume, transparent output comes from ChatGPT Image 2.5's background field.
Qwen-Image 2.1 is 7B: self-host it or call hosted qwen-image on Sume
Qwen-Image 2.1 is a 7B model under the Qwen Research License. Sume hosts qwen-image and qwen-image-max, not 2.1; hosted costs $0.025 or $0.094 per image.
Seedream 5.0 Flash takes 14 references; Sume seedream-5-lite takes 10
Seedream 5.0 Flash lists up to 14 input references on OpenRouter. Sume's catalog caps references at 10 and lists seedream-5-lite, not Flash.
Seedance output size in pixels: what 720p 9:16 or 1080p 21:9 means
Pixel dimensions Sume uses for Seedance by resolution and aspect ratio: 720p 16:9 is 1280x720, 1080p 21:9 is 2205x945, 480p 9:16 is 496x864.
AI gift basket product photo from six item photos: 16 references
GPT Image 2.5 on Sume takes up to 16 reference images, so six item photos can go into one gift basket scene; and the billed price is in usage.cost.
AI book cover mockup on Nano Banana Pro: 4:5 and four resolutions
Nano Banana Pro lists at $0.15 at 1K and default quality and bills $0.1875 on Sume; it takes 0.5K, 1K, 2K or 4K and 4:5, so read usage.cost for larger tiers.
Ceramic mug product photos: Nano Banana Pro hero, draft on Banana 2
Draft a ceramic mug scene on Nano Banana 2 at $0.10, then render the hero on Nano Banana Pro at $0.1875 in 4:5; six mugs cost $1.725 in total.
AI flower bouquet photos for a florist: Seedream 4.5 reference edit
A florist can keep a real arrangement and change only the backdrop on Seedream 4.5: $0.04 list, $0.05 billed on Sume, 4:5 ratio, up to 10 references.
Seedance 2.0 Mini, Fast, Standard or 2.5: which tier for ad drafts
On Sume the Seedance tiers price per 1,000 video tokens: Mini is half of Standard, Fast is 80%, and 2.5 is about 53% above Standard. Mini is the draft tier.
Christmas tree lot flyer art with GPT Image 2.5: quote every line
Tree lot flyer on GPT Image 2.5: custom 1536x2048 is legal, quality defaults to high, and usage.cost shows the billed price.
2K or 4K AI video on Sume: H3 upscale or Omni 4K, and what each costs
For 4K on Sume, H3 is $0.20 a second as an upscale of its 768p output and Omni is $0.375. A 10-second 4K clip is $2.00 on H3 and $3.75 on Omni.
AI photo booth strip from one selfie: a 1:4 Nano Banana 2 edit
A four-frame photo strip fits the 1:4 ratio that Nano Banana 2 lists on Sume; one selfie reference and one call cost $0.10 billed per strip.
AI greeting card front art on Qwen Image Max: ratio and price check
Qwen Image Max on Sume gives card-front art in 3:4, 4:5 or 2:3 portrait; read the catalog's n range and the billed usage.cost before pricing a card line.
AI tattoo flash sheet: 12 designs in one Ideogram V3 image, sliced
One Ideogram V3 image holds a 3x4 sheet of 12 flash designs for $0.075 billed; slice it into 12 files in Pillow instead of paying for 12 calls ($0.90).
AI cocktail menu photos: four 4:5 takes per drink on Sume
Eight cocktails at four takes each is 32 images: $3.20 on Nano Banana 2 at $0.10 each, 4:5 ratio, one call per drink with n=4.
Which aspect_ratio values do Sume video models list: 9:16 and 16:9
Sume docs list 9:16 and 16:9 for Auto and Gemini Omni Flash 1.1. Other models vary, so read capabilities from the catalog. What to send for TikTok or Reels.
9:16 cover still: 1080x1920 fails GPT Image 2.5, 1088x1920 passes
GPT Image 2.5 needs edges in multiples of 16, 655,360 to 8,294,400 pixels, and a ratio of 3:1 or less. 1088x1920 and 1152x2048 pass; 576x1024 is too small.
Seedance 2.5 and Wan 3.0 claim 30 s: what Sume's ids accept
Both vendors list 30-second generations. Sume's seedance-2.5 accepts 4-30 s and wan-3.0 2-30 s. The lower bounds, resolutions and reference limits differ.
Wan 3.0 text in 12 languages: what Sume sends and what it bills
Alibaba says Wan 3.0 renders long text in 12 languages. On Sume you steer it through the prompt: no language field, 480p $0.0625 to 1080p $0.25 a second.
H3 or H3 Max? Pick the Sume row by job, with per-clip prices
Both read references and make stereo audio. H3 is $0.075 per second at 768p, H3 Max $0.10 with a 1080p row. A job-by-job guide with 10-second prices.
Ideogram 4.5 specs: 25 MB and 1:6 to 6:1 vs Sume's 1K/2K, 1:3 to 3:1
A partner page lists 2K, 25 MB inputs and 1:6 to 6:1 for Ideogram 4.5. Sume's catalog allows 1K or 2K and 1:3 to 3:1. Which limits apply to your call?
Ideogram 4.5 open weights are "soon": what to build on the API now
Ideogram says 4.5 open weights are coming. Until then it is an API call: the Sume id, $37.50 to $275 per 1,000 edits, and how to keep a later switch cheap.
Wan 3.0 box-selection editing: Sume offers Omni video_url edit instead
Alibaba lists box-selection editing for Wan 3.0. Sume documents no box field for it; its edit route is Gemini Omni Flash 1.1 video_url at $0.125 a second.