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
Seedance 2.5, Wan 3.0, Kling 4.0: which model ids Video Router accepts
Video Router lists 12 catalog ids plus three Auto spellings. seedance-2.5 and wan-3.0 are there; Kling 4.0 is not, and kling-3 is the Kling id. Limits per id.
Seedream 5 Lite product photo background swap: what Sume offers
Swap a product photo's background with Seedream 5.0 Lite: ByteDance's edit page facts, and the Sume id, ratios and n range you can call today.
Reuse a Sora 2 prompt on Gemini Omni Flash: what to rewrite
Moving a Sora 2 prompt to Gemini Omni Flash 1.1 on Sume: keep the camera and light words, rewrite dialogue blocks, add a single-shot line, fit 3 to 10 seconds.
Sora's short-clip dialogue rule, tested on Gemini Omni Flash
OpenAI's Sora guide fits 1-2 exchanges in 4 seconds. Carry that rule to Omni Flash clips on Sume and check the words with a video inspect transcript.
Split a 16 or 20 second Sora shot into Omni clips and join them
A Sora shot of 16 or 20 seconds becomes two or three Gemini Omni Flash clips of 10 seconds or less on Sume, joined with Timeline 1.0. How to cut the beats.
Which Sume image model lists the most aspect ratios?
Ideogram V3 and 4.5 list 15 ratios; Nano Banana 2 lists 14 plus auto, including 8:1 and 1:8. Ratio counts for all 19 Sume image rows, with what to pick.
Is Gemini 3.8 Flash on Sume? An agent LLM, not a media model
Gemini 3.8 Flash is on Sume as an agent LLM that the auto setting uses before a GPT hop. It does not make video or images; those are separate catalog ids.
Design an episode's opening frame with Grok Imagine, then animate it
Make the first frame with x-ai/grok-image (9:16, about $0.025 a still on Sume), approve it, then pass it as first_frame to a video model. Costs and code.
Which Sume image models can't edit? Five text-to-image-only rows
Five Sume image rows take text prompts only and reject image_urls: Soul, Imagen 4 Fast and Ultra, Recraft V4 and Qwen Image Max. Prices and what to use instead.
AI video ratios on Sume: Omni Flash is 16:9 or 9:16, Kling adds 1:1
Gemini Omni Flash 1.1 on Sume takes only 16:9 and 9:16; Kling 3 adds 1:1. Which other Sume video models list more ratios, and which list none.
Cold open with a recurring cast: Gemini Omni Flash with 10 image refs
Gemini Omni Flash 1.1 on Sume takes up to 10 image refs and 3 video refs of 3 s each for a 3 to 10 s vertical cold open with synced audio. Request and limits.
First frame plus style sheet in one video call: frame_images wins
On Sume /v1/videos, frame_images plus input_references runs as image-to-video, so plan on your style sheet being set aside. Split the call and pin the opening.
Gemini 3.8 Flash is stable: keep model ids in config
When a vendor ships a new Flash model, a hard-coded id ages. Read Sume ids from GET /v1/catalog and treat 404 model_not_found as a signal, not a retry.
16:10, 10:16, 3:1 and 1:3 image ratios: only Ideogram on Sume
Only the two Ideogram models in the Sume catalog list 16:10, 10:16, 3:1 and 1:3. Ideogram V3 bills $0.075 per image; read the list before sending.
Ideogram 4.5 ten-edit chain: cost and drift on Sume
Ten sequential edits on Ideogram 4.5 cost $0.75 at medium quality on Sume. Edit from the original, not the last result, to avoid compounding drift.
MAI-Voice-2.1: 14 serving regions, public preview, no SLA
MAI-Voice-2.1 is in public preview with no SLA and 14 serving regions. A pre-launch checklist, plus how an async TTS job changes the risk.
MAI-Voice-2.1 voice ID format and which voices have styles
A MAI-Voice voice is named like en-US-Harper:MAI-Voice-2.1-Flash. Style lists differ by voice: some have 20 styles, some only neutral.
Mercury Voice p95 750 ms: one turn in twenty is slower
Inception reports a 320 ms median and 750 ms p95 for Mercury Voice. At p95, about one turn in 20 is slower. What that means over a 10-turn call.
Mercury Voice is enterprise-only: pricing and what to ask sales
Mercury Voice is GA for enterprise customers only. List $0.40/$1.50 per million tokens, launch price $0.20/$0.75. Questions to ask before you commit.
Sume TTS Router: five Sonic ids, same price as TTS 1.0
Sume TTS Router needs a model: one of five Sonic ids. It uses the same $0.0475 per 1,000 characters as TTS 1.0. What the model field changes.
Sume Music Router: job.model vs routed_model, which engine ran
On Sume Music Router, job.model echoes the id you sent and request.routed_model names the engine that ran. Ids, price and the Music 1.0 route.
Pocket TTS speed: 6x on an M4 CPU, 2.3-2.5x on a VM CPU
Kyutai Pocket TTS runs about 6x real time on an M4 CPU and 2.3-2.5x on a 4-vCPU cloud VM. It cannot add silence for pauses. What to plan for.
MAI-Voice instant cloning: gated access and a 5-60 second clip
MAI-Voice cloning needs approval through Microsoft's Limited Access Review and a 5-60 second consented clip. What that means for your launch plan.
Sume Image API n: 1 to 10 per call, lower per-model ceilings
The n parameter asks for 1 to 10 images in one POST /v1/images, but each model's catalog entry can set a lower ceiling. Cost is per image times n.
NVIDIA VideoFDB: what Griffin Lite's 3.73 means for avatar buyers
VideoFDB scores perception on 237 call clips across 11 nonverbal dynamics. Griffin Lite led full-duplex systems at 3.73 against a human 4.20.
MAI-Voice-2.1 vs Flash: Microsoft's best-for column, read closely
Microsoft lists MAI-Voice-2.1 for long-form narration and Flash for real-time agents, and notes the full model favors naturalness over latency.
Which Lyria engine made my Sume track? Read routed_model
sume/music-auto picks Lyria 3.5 today. The job echoes the id you sent and names the engine in request.routed_model. Pin lyria-3-pro by id. All cost $0.125.
Which of Sume's nine aspect ratios work on Reels, Shorts, TikTok ads
Sume lists nine video aspect ratios. Checked against Instagram's 1.91:1 to 9:16 Reel range, YouTube's square-or-vertical Shorts rule and TikTok's ad ratios.
Which Sume video models clear Instagram's 720-pixel Reel minimum
Instagram lists a 720-pixel minimum for Reels. The Sume catalog shows which video models offer 720p or above, their clip lengths, and which fall short.
AI Product Video Looks the Wrong Size in Hand: Scale Reference Fix
A bottle that looks twice its real size in a generated hand is a scale error. Add a size reference photo, state the dimensions, and verify with stills.
Product in Hand Video Without a Hand Model: Reference-to-Video QC
Show your product held in a hand with no shoot: give a reference-to-video model product photos, then check four frames for label, scale and fingers.
H3 Max Recast file types: mp4, mov, webm, m4v or GIF?
fal lists mp4, mov, webm, m4v and gif as Recast inputs. What to check before sending a clip to Sume: a public HTTPS URL, a rounded-up duration, one probe.
Recast a clip with more than four people: the four-photo cap
Sume accepts 1 to 4 reference photos for H3 Max Recast, one per new person. What happens with a group shot, and three ways to work within the cap.
Which Sume image model for a photo edit: mask, references or pixels?
Pick between ChatGPT Image 2.5, Ideogram 4.5 and Nano Banana 2 for a photo edit on Sume by what the edit needs: a mask, many references, or untouched pixels.
Why sume/auto never picks H3 Max Recast, and what to send
A person swap needs a source video and photos. Sume's Auto router never routes to Recast, so you name h3-max-recast yourself. Here is the request.
gemini-omni-1.1-flash vs gemini-omni-flash-1.1: which id goes where
Google writes the model id gemini-omni-1.1-flash; Sume's catalog id is gemini-omni-flash-1.1. A mapping table, the old preview id, and how to avoid a typo.
GPT Image 2.5 mask_url edit: change one region, keep the rest
GPT Image 2.5 on Sume accepts a public mask_url alongside input_references. Build a mask with Pillow, host it, and edit only one region of a photo.
MAI-Image-2.6 output cap is 2,359,296 pixels; Sume sizes differ
MAI-Image-2.6 sets a 2,359,296-pixel ceiling and a 768-pixel minimum edge. Sume sets size per model with tiers, ratios and, for GPT models, custom pixels.
MAI-Image-2.6 edits take 5 references; Sume takes 10 or 16
MAI-Image-2.6 in Foundry accepts up to five JPEG or PNG reference images per edit. On Sume, input_references tops out at 10, or 16 on GPT Image 2.5.
MAI-Image-2.6 web_grounding flag: what Sume has instead
MAI-Image-2.6 can pull Bing results into an image when web_grounding is on. Sume has no such flag; here is how to pass current facts in the prompt instead.
MiniMax H3 Max Recast API: swap people in a video, fal price vs Sume
H3 Max Recast swaps people in a source video for reference photos, keeping motion, cuts and audio. fal lists $0.30 a second at 768p; what Sume accepts.
Swap the product in a UGC clip with Gemini Omni Flash 1.1 video edit
Keep a UGC clip that works and change only the product: Sume's video router edit mode for gemini-omni-flash-1.1 takes a video_url and a one-line prompt.
Which AI video model for a 15-second single take on Sume?
Kling 3.0, Wan 3.0, MiniMax H3 and Seedance 2 reach 15 seconds on Sume; Auto and Grok stop at 10. Seedance 2.5 and Wan 3.0 go to 30. Table of limits.
Does H3 Max Recast keep the original audio? What to check on Sume
fal says H3 Max Recast preserves the source audio. Sume rejects generate_audio and audio references on it. Confirm a result has sound with video inspect.
Grok Imagine video in Sume: why it needs a start-frame image
Sume's Grok Imagine entry is image-to-video only: it blocks a submit without a start frame, tops out at 10 seconds, and sends no audio or aspect ratio.
H3 Max 3D to Video: previs to photoreal on fal, not on Sume
fal's H3 Max 3D-to-Video turns a blockout render into photoreal video for $0.50 a request plus per second. Sume lists no such endpoint; what it offers instead.
H3 Max Insert-Video: add a scene mid-clip on fal, and on Sume
fal's H3 Max Insert-Video adds 5 to 13 s to a source up to 60 s, billed on the new seconds only. Sume lists no such row; here is a trim, generate, join route.
H3 Max Recast prompt: optional, and what Sume does with one
On Sume, H3 Max Recast runs without a prompt: the 1-4 photos say who to swap in. A prompt is allowed up to 2000 characters. Request body and limits below.