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
Kling 3.0 4K in Sume's Videos panel: what actually gets submitted
The panel lists 4K for Kling 3.0 but its live submit maps 4K down to 1080p. Where 4K does exist on Sume's API, and how to check the model before you pay.
MiniMax H3 camera prompts: lens, movement, exposure wording
fal's H3 prompting guide says H3 reads film vocabulary: lens, rack focus, handheld, grain. Wording that works as a prompt and a Sume request that sends it.
MiniMax H3 sound design prompts: direct the audio like the picture
fal's H3 guide says to direct audio as deliberately as picture: name sonic elements, not 'music'. What it looks like in a Sume request, and what you can't set.
Sume Auto video has no 1:1 or 21:9: which model to pin instead
Auto in Sume's Videos panel offers 720p or 1080p, 16:9 or 9:16 and 3 to 10 seconds. For square, ultrawide or longer clips, pin Wan 3.0, H3 or Kling 3.0.
Reference limits per Sume video model: images, clips, audio
How many reference images, videos and audio clips each Sume video model takes on /v1/videos: Seedance 12 total, Wan 10/5/5, H3 9/3/3, Omni 10 and 3.
Which AI video model gives 1080p on Sume, and which stop at 768p?
Sume lists 1080p for Seedance, Wan 3.0, Kling 3.0 and Auto; MiniMax H3 is native 480p or 768p in the panel. Resolution table by model, with the API check.
flux-2-pro-preview vs flux-2-pro: which id does Sume send?
BFL has flux-2-pro-preview (latest) and flux-2-pro (fixed snapshot). Sume lists black-forest-labs/flux.2-pro and does not let you pick the BFL endpoint.
Gemini Omni audio reference: unsupported; Sume models that take audio
Google says Omni's API doesn't accept uploaded audio references. Sume's Omni row has no reference_audio_urls; Seedance 2.x, Wan 3.0 and MiniMax H3 honor audio.
Gemini Omni extend: no new dialogue on an uploaded talking clip
Google says you can't extend an uploaded Omni clip where someone talks to add dialogue; extension only appends, to clips up to 10 s. Sume lists no extend mode.
Gemini Omni with several videos: 3 references, no cross-video use
Gemini Omni takes up to 3 reference clips of 3 s each, yet Google warns that reasoning across several videos may degrade output. Sume: one video_url source.
Gemini Omni REST: output_video is SDK-only, read the steps array
Calling Gemini Omni over REST? interaction.output_video is SDK-only. Read the base64 video from the model_output step; on Sume you get a media URL instead.
Gemini Omni [# Sources] and [# References] tags vs Sume fields
Gemini Omni binds media to roles with tags like <FIRST_FRAME> and [# Sources ...]. Sume uses request fields instead: image_url, end_image_url, reference lists.
Gemini Omni and Veo in Korean: only English is fully supported
Google's Omni and Veo pages say English is fully supported; other languages aren't evaluated. For Korean prompts, describe in English and quote on-screen text.
Gemini Omni can't use a YouTube link as its source video
Google lists YouTube videos as an unsupported media source for Gemini Omni Flash. On Sume, video_url is a URI field; use a direct link to the clip file.
GPT Image 2.5 character drift: what OpenAI says and what to send
OpenAI's image guide says GPT Image may struggle to keep a recurring character the same across generations. What to send on Sume to reduce drift.
Kling motion control keep_original_sound: get a silent clip
keep_original_sound on Sume's Kling 3.0 Motion Control defaults to true, so the driving video's audio rides along. Send false for a silent clip.
LTX-2.3 LoRAs on LTX-2.5: most run unchanged; Sume has no LoRA field
Lightricks says the large majority of LTX-2.3 LoRAs and IC-LoRAs run on LTX-2.5 unchanged, with a few exceptions to test. Sume has no LoRA field.
LTX-2.5 distilled or dev transformer: which file to download
Download the distilled bf16 transformer to generate (fixed 8-step schedule, CFG 1) and the dev one to train. int8 files are ComfyUI-only. Sume lists no LTX id.
LTX-2.5 duration predictor: optional, and Sume sets seconds
LTX-2.5 has an optional duration predictor that sets the frame count from the prompt. Sume has no such mode: you send duration in whole seconds.
Is LTX-2.5 gated on Hugging Face? Agree, log in, then download
Yes: the LTX-2.5 repository is gated. Agree to share your contact details on the model page, run hf auth login, then hf download. It is free under $10M revenue.
Where are the lyrics in a Lyria 3.5 result? Gemini vs Sume
Google returns Lyria 3.5 lyrics and song structure as text beside the audio. Sume puts model-reported lyrics or a section map in result.lyrics when present.
Nano Banana negative prompt: describe what you want instead
Google's Gemini image guide says to write semantic negative prompts: describe an empty street, not 'no cars'. Sume's image request has no negative field.
Nano Banana prompt languages: Korean and Japanese, per Google
Google lists the languages Gemini image models work best in, including ko-KR and ja-JP. What that means for a Korean or Japanese prompt sent through Sume.
Nano Banana text in images: write the copy first, then render
Google's tip for text in Gemini images: settle the wording first, then ask for the image. How to do that with one Sume request and a short checklist.
Seedance 2.5 concert prompt uses 18 images: fitting Sume's 9
Seed's Seedance 2.5 concert example uses 18 reference images. Sume's Video Router takes at most 9 for seedance-2.5, so merge groups into sheets.
Seedance 2.5 duration 3 seconds: 400 error, the minimum is 4
Seedance 2.5 on Sume takes 4 to 30 seconds. A 3-second request returns 400; a 31-second one is refused too. Here are the messages and the shorter-clip options.
Seedance 2.5 native 4K: the press claim and what Sume accepts
A June report said Seedance 2.5 makes native 4K. Sume lists 480p, 720p and 1080p for seedance-2.5 and returns 400 for a 4K request. What to send instead.
Seedance 2.5 one-take prompt: one continuous shot, 30 seconds
Seed's Seedance 2.5 one-take prompts ask for a single continuous take with no cuts. Write the camera path in beats and send it to Sume as a 30-second job.
Seedance 2.5 adds subtitles or music I did not ask for: fixes
Seed says Seedance 2.5 minimises stray subtitles and music. On Sume, generate_audio false drops the audio track; subtitles need prompt wording.
Veo seed doesn't make output repeatable; Sume rejects seed
Google says Veo's seed only slightly improves determinism. Sume's v1 video models report seed false and reject the field; keep the output file to get a repeat.
Wan 2.2 weights: Apache 2.0, 5-second 480p/720p; Sume hosts wan-3.0
Wan 2.2 T2V-A14B is Apache 2.0 with open weights and makes 5-second 480P or 720P clips. Sume does not host Wan 2.2; its Wan row is wan-3.0, 2 to 30 seconds.
Can GPT-6 Sol, Luna or Astra generate images or video?
No. OpenAI lists GPT-6 Astra, Sol and Luna as text-output models with image input. Where image and video generation live instead, and how Sume splits it.
Does sume/auto pick Seedance? Sume does not say which model ran
Sume says sume/auto echoes sume/auto in the response and never discloses the family that served the request. To get Seedance, pin the id.
FLUX.2 takes 8 references by API, 10 in Playground: Sume's range
Black Forest Labs says FLUX.2 pro, max and flex take up to 8 references by API (10 in its Playground), klein 4. Sume's input_references range is read per model.
gpt-image-2.5 smallest custom size: the 655,360-pixel floor
gpt-image-2.5 on Sume rejects images under 655,360 pixels. Why 512x512 and 768x768 fail, which small sizes pass, and the other edge rules to check first.
Nano Banana reference limits by model: objects, characters, style
Google splits Nano Banana references by model: Nano Banana 2 takes 10 objects, 4 characters, 3 style images; Pro takes 6, 5, 3. Sume uses one list.
Qwen-Image 2.0 Pro is Alibaba's pick: which Qwen ids does Sume list?
Alibaba recommends qwen-image-2.0-pro. Sume's image catalog lists qwen/qwen-image and qwen/qwen-image-max in the repo; check the live catalog for more.
Qwen-Image negative_prompt 500 characters vs a Sume image request
Alibaba's Qwen-Image API takes a negative_prompt up to 500 characters. Sume's image request table has no such field, so rewrite exclusions into the prompt.
Seedance 2.5 on Jimeng or Doubao Pro: where the API fits
ByteDance launched Seedance 2.5 on Jimeng AI and Doubao Pro. If you need it from code, here is what Sume lists and what the vendor page says.
Seedance 2.5 timestamp-level editing: no timestamp field on Sume
ByteDance describes timestamp-level control in Seedance 2.5. Sume documents no timestamp parameter on /v1/videos. Here is what you can send instead.
Wan 3.0 allows 20,000 prompt characters, MiniMax H3 7,000
Alibaba allows 20,000 characters in a Wan 3.0 prompt; MiniMax caps each H3 text item at 7,000. How to write one prompt that fits both on Sume.
Wan 3.0 input video plus output capped at 30 seconds: Sume plan
Alibaba caps Wan 3.0 input video plus output at 30 seconds. What that means for reference_video_urls and duration when you call wan-3.0 on Sume.
Lyria 3.5 is single-turn and varies per call: keep the artifact
Google says Lyria generation is single-turn, not iteratively editable and varies between calls. Why to save every Sume Music artifact you like.
ByteDance MPA copyright pact: what Seedance users should check
ByteDance signed a copyright agreement with the MPA on Aug 17, 2026 covering Seedance. What was disclosed, what was not, and how to handle a refused request.
Claude structured outputs allOf limits vs Sume output_schema
Anthropic supports allOf with limitations in structured outputs. Sume's output_schema rejects allOf and oneOf and accepts anyOf. How to port a schema.
Claude structured outputs minItems only 0 or 1 vs Sume
Anthropic's structured outputs accept minItems of 0 or 1 only. Sume's output_schema accepts minItems and maxItems, enforced on the projection. The differences.
GPT Image 2.5 edit drifts after a few turns: repeat the preserve list
OpenAI says to repeat the preserve list on each iteration to reduce drift. How to run a one-change-per-call edit chain on Sume, which keeps no chat memory.
GPT Image 2.5 repeats the headline: say how many times it appears
Duplicate text in a GPT Image 2.5 result? OpenAI's guide says to quote the copy and state how often it appears. The prompt, plus n and billing on Sume.