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
GPT Image mask edit changed pixels outside the mask: what OpenAI says
OpenAI says GPT Image masks are guidance and may not follow the exact shape. Why a mask_url edit leaks, and how to keep the rest of the image on Sume.
GPT Image 2.5 misspells a brand name: spell it letter by letter
OpenAI's prompting docs say to quote the text and spell tricky words letter by letter. The prompt to write, and how to check the result on Sume's image API.
LTX-2.5 native multishot: one prompt, or clips joined on Sume
LTX-2.5 adds native multishot: connected scenes in one pass, consistent characters. Sume has no LTX; its route is separate clips joined on Timeline 1.0.
Lyria 3.5 in another language: prompt in it, then check the song
Google says Lyria 3.5 makes music in other languages when you prompt in that language. On Sume, write the brief and lyrics in it, then check the result.
Meta Muse Video: preview only, so what can you use for Reels?
Meta's July 7, 2026 post previews Muse Video with native audio and says it is coming soon. No API details yet. How to check what Sume lists for 9:16 video.
Seed Audio 1.0 on OpenRouter: text to speech, and what Sume offers
ByteDance's Seed Audio 1.0 is listed on OpenRouter as text-to-speech with raw MP3 or PCM output. Sume's TTS Router is Cartesia Sonic only today.
Seedance 2.5 on Dreamina: age 16+ and countries, what applies
Dreamina says Seedance 2.5 is for subscribers over 16 and names rollout regions. What the pages say, and where a developer account differs.
SynthID audio and MP3: does the mark survive Sume's 128 kbps detach?
Google says SynthID audio marks resist noise, MP3 compression and speed changes. Sume audio detach can output 128 kbps MP3; only some sources carry a mark.
SynthID text watermark: a Gemini script in a Sume avatar video
Google says SynthID marks Gemini app text. If it becomes a Sume avatar script, the video is a new render, and no docs say the mark carries over.
Wan2.2 S2V-14B speech-to-video: open weights vs Sume audio options
Wan2.2-S2V-14B is an Apache 2.0 audio-driven model at 480P and 720P. On Sume: an audio reference on Wan 3.0, or a script-driven Avatar video.
Zalando designer minimum 1800x2600: request 1808x2608 on Sume
Zalando designer brands need at least 1800x2600 px in 1:1.44 JPEG. GPT custom sizes need multiples of 16, so ask Sume for 1808x2608, which clears both edges.
Dialogue in Korean, Japanese or Spanish: Kling 3.0 vs Seedance 2.5
Which spoken languages Kling 3.0 Omni and Seedance 2.5 list, what Sume's generate_audio flag does, and how to test a Korean line before a full render.
Sonnet 5 retired on Sume: requests move to Sonnet 5.5, ids explained
Sume moves every request for Claude Sonnet 5 onto Sonnet 5.5 while Anthropic still lists Sonnet 5 as legacy. The three id spellings and what stays on Sonnet 5.
DeepSeek V4.1 Flash: the model Sume Auto starts on, and what it takes
Sume's Auto agent model reserves and starts on DeepSeek V4.1 Flash. DeepSeek's page lists a 1M window, 384K output and vision. How the served model settles.
Gemini image multi-turn editing: how to loop edits on Sume
Gemini's image models edit across chat turns. Sume's Image API has no chat session, so you pass the last result URL back as a reference each round.
gpt-5.6-terra and gpt-5.6-luna on Sume: which still run, as what
Sume retired GPT 5.6 from every picker. Terra has no successor and runs as itself, Luna moves to GPT-6 Luna only where that is admitted, Sol moves to GPT-6 Sol.
GPT-6.1 Sol effort on Sume: picker offers Low, Medium, High, plus Fast
OpenAI lists five effort values for GPT-6.1 Sol. Sume's picker exposes three, keeps effort off the model id, and prices Fast as an opt-in. What that means.
GPT Image 2.5 text in the wrong place: fix it with a mask edit
OpenAI says GPT Image 2.5 can still misplace text. Mask only the text zone and re-render that region through Sume's mask_url edit instead of the whole image.
grok-4.5 as a Sume model id runs GPT-6 Sol, not Grok: the alias list
Sume treats grok-4.5, composer-2.5, composer-2.5-fast and sume-1.0-fast as old Cursor-era ids that run GPT-6 Sol. How Grok 4.7 is gated and how grok-4.6 moves.
HunyuanVideo 1.5 step-distilled: 8 or 12 steps for 480p I2V
HunyuanVideo-1.5 ships a step-distilled 480p image-to-video model that needs 8 or 12 steps. What the README claims, the model sizes, and when hosted is easier.
Is Wan 2.7 open source? The official Wan-AI org lists Wan 2.2
Wan-AI's Hugging Face page lists Wan 2.2 weights and no 2.5, 2.6, 2.7 or 3.0 weights. Here is how to check, and which hosted route Sume lists for Wan.
LTX-2.5 quantization: fp8-cast, fp8-scaled-mm or NVFP4?
LTX-2.5's repo offers fp8-cast with bf16 checkpoints and fp8-scaled-mm on Hopper; the Hugging Face card lists NVFP4 and int8. Which flag fits which GPU.
MiniMax H3 FL2VA vs Ref2VA: reference requests need own server
MiniMax splits H3 into FL2VA (text, first/last frame) and Ref2VA (images, video, audio references) checkpoints. Which tasks go where, and the Sume model ids.
Mochi 1 LoRA on one H100 vs reference images on a hosted API
Mochi 1's trainer fine-tunes a LoRA on one H100 or A100 80GB. Sume has no LoRA field, so here is when to train and when to send reference images.
Opus 5 retired on Sume: why requests and threads show Opus 5.5
Sume moves Claude Opus 5 requests onto Opus 5.5, listed on every catalog. What Anthropic's page says, what history keeps, and how to check which Opus ran.
Same voice across AI video clips: Kling Omni voice vs Seedance audio
Kling 3.0 Omni binds a voice from a 5-30 second sample; Seedance takes up to 3 audio references. What each page says and what Sume lets you send.
Seedance prompt for sound: dialogue, ambience and music lines
Seedance generates voices, effects and music with the picture. How to write each as its own prompt line, what Sume's generate_audio flag does, and its limits.
Wan 2.2 A14B explained: 27B total, 14B active, 80GB GPU
Wan 2.2's A14B models use two experts, one for noisy early steps and one for late detail. What that means for GPU memory, and when a hosted route is simpler.
Which AI image model edits a photo best? Reference limits compared
Reference-image limits for edits on GPT Image 2.5, Nano Banana, Grok Imagine Image 2.0 and FLUX, from each vendor's page, and how to call them on Sume.
Cached input price for agent LLMs: GPT-6.1 Sol to DeepSeek
Cached input per million tokens for GPT-6.1 Sol, Opus 5.5, Sonnet 5.5, Grok 4.7 and DeepSeek Flash, and why it sets the cost of a long video agent thread.
Context window and max output for agent LLMs, compared
Context and output limits for GPT-6.1 Sol, Opus 5.5, Sonnet 5.5, Grok 4.7 and DeepSeek Flash from vendor pages, and what they mean for a long video thread.
POST /v1/avatar-1.0/fabric is test-only: use veed/fabric-1.0 instead
The experimental avatar-1.0/fabric route is a temporary comparison endpoint that may be removed. How it differs from image-to-video and what to build on.
ChatGPT Image 2 rejects quality xhigh: only 2.5 takes xhigh and max
Sending quality xhigh or max to openai/gpt-image-2 on Sume returns 400. The 2.5 ids accept auto, low, medium, high, xhigh and max. Table and a safe fallback.
Opus 5.5 and Sonnet 5.5 retirement dates for saved Formats
Anthropic lists Opus 5.5 and Sonnet 5.5 as not retiring before September 2027. What that means for saved Sume Formats, model ids and the receipt's model field.
Opus 5.5 or Fable 5.1 for a video agent: when to step up
Anthropic says start with Opus 5.5 and move to Fable 5.1 only when evals fall short. Price, latency, effort and a one-turn cost example for a video agent.
DeepSeek legacy names now run V4.1 Flash: what Sume offers
DeepSeek still accepts deepseek-v4-flash and the vision-exp name but serves V4.1 Flash. Which DeepSeek rows Sume lists, and why there is no v4-pro row.
Grok 4.7 doubles price at 200k tokens: long video agent threads
xAI bills a Grok 4.7 request at the higher rate for every token once the prompt reaches 200k. What that does to a long video agent thread, with the arithmetic.
HunyuanVideo 1.5 super-resolution to 1080p vs hosted 1080p
HunyuanVideo 1.5 makes 480p or 720p clips of 121 frames and a super-resolution network lifts them to 1080p. What a hosted 1080p second costs on Sume.
Image resolution tiers on Sume: Nano Banana 512 to 4K, 400 elsewhere
Only Nano Banana and Soul take resolution on Sume. GPT Image, Seedream, Grok and FLUX.2 return 400. What size accepts, and where pixels go instead.
LTX-2.5 Ingredients LoRA: a reference sheet for consistent clips
LTX-2.5 Ingredients conditions a 121-frame clip on a character, prop and location sheet. How the sheet works and how Sume input references differ.
LTX-2.5 Layout-To-Render: Blender playblast to a finished shot
LTX-2.5 Layout-To-Render keeps a 3D playblast's camera path and layout and restyles it from a reference image. What Sume's video references can and cannot do.
LTX-2.5 Pixel Spatial Upscaler: draft at 280p, then upscale 2x
The LTX-2.5 Pixel Spatial Upscaler re-renders a low-res draft 2x with invented detail. How the draft workflow works and how Sume's draft-then-final differs.
LTX-2.5 Refine Details: 22 minutes per clip on an RTX PRO 6000
The LTX-2.5 Refine Details LoRA rebuilds texture up to 4K but takes about 22 minutes for 121 frames on an RTX PRO 6000. A hosted upscale compared.
LTX-2.5 Slow-Motion Control: speed 0.025 to 1.0, 40x slower
The LTX-2.5 Slow-Motion Control LoRA sets motion speed from 0.025 (40x slow) to 1.0 while playback stays 24 fps. Sume's video API has no speed field.
LTX-2.5 Water Simulation LoRA: add rain or a flood to a clip
LTX-2.5 Water Simulation adds water to a dry reference clip at native 1920x1088. How its prompt works, and the hosted edit route on Sume instead.
Luma API camera concepts: 34 named moves vs Sume's prompt-only camera
Luma's API lists 34 named camera concepts you can layer, such as orbit_left and dolly_zoom. Sume has no camera field, so here is how to ask for the same moves.
Luma Photon API: 4 image refs and a character ref vs Sume references
Luma's photon-1 image API takes up to 4 image references, a style reference and a 4-image character reference, each with a weight. Here is the Sume equivalent.
mask_url and background on Nano Banana, Seedream, FLUX, Grok: 400
On Sume only the two GPT Image 2.5 ids accept mask_url and background. Other image models return 400 unsupported_parameter. What to use for a local edit.