Luma Ray 3.2 features and the Sume endpoint for each one
Sume does not run Ray 3.2. Match its keyframes, reframe, face tracking and 20 s clips to the Sume surfaces that exist, with docs-verified limits.

Sume does not run Luma Ray 3.2: it is not in the Sume video catalog, so you cannot call it here. If you came for what Ray 3.2 does, the closest Sume equivalents are first and last frames on catalog models, clip chaining, the Timeline API for multi-shot assembly, and video filter and trim for reframing and cuts.
This page maps each Ray 3.2 feature that Luma lists to the Sume surface that covers part of it, and says where there is no match. Ray 3.2 facts come from Luma's launch page; Sume facts come from the Sume docs.
What Luma lists for Ray 3.2
Luma's launch page for Ray 3.2 (dated 2026-06-09) lists up to 16 keyframes per clip, 1080p, clips up to 20 seconds, native HDR with 16-bit EXR export, facial performance tracking for up to 8 faces, a Reframe feature, and a full control surface through its API.
None of that is a statement about Sume. It is the checklist we use below.
Feature map
| Ray 3.2 feature | Closest Sume surface | Gap |
|---|---|---|
| Up to 16 keyframes per clip | frame_images with first_frame and last_frame on /v1/videos; chain clips for more | Two frame types per request, not 16 |
| Up to 20 s per clip | seedance-2.5 accepts 4-30 s, wan-3.0 2-30 s; most other ids max 15 s | Different per model; read capabilities |
| 1080p | seedance-2.5 offers 480p/720p/1080p; Omni Flash 1.1 goes to 4K | Varies by model |
| HDR, 16-bit EXR | Not documented. Sume returns MP4 | No HDR or EXR claim in the docs |
| Facial tracking, 8 faces | h3-max-recast replaces 1-4 people in a source video | Swap, not tracking data |
| Reframe | Video filter crop plus Timeline fit modes | Crop is geometric, not subject-aware |
| API control surface | POST /v1/videos with Idempotency-Key, polling or callback_url | Sume pins one catalog id per request |
Keyframes: two frame types, plus chaining
On POST /v1/videos, the frame_images array takes entries with a frame_type of first_frame or last_frame. Whether a model accepts them is listed in supported_frame_images on GET /v1/videos/models. That gives you a start and an end image for one clip.
To go beyond two anchors, generate clip one, extract its final still with video frames, and send that still as the first_frame of clip two. Each link is a separate job, so every hop is a separate bill.
Length and resolution
Limits differ by model. The docs state that seedance-2.5 accepts 4-30 seconds at 480p, 720p and 1080p, wan-3.0 accepts 2-30 seconds, and each other catalog model has a ceiling of 15 seconds. Gemini Omni Flash 1.1 (gemini-omni-flash-1.1) is 3-10 seconds at 360p, 720p, 1080p or 4K. So a 20-second single take exists on Sume only on the 30-second models.
Reframe, assembly and edits
For a 16:9 to 9:16 change, video filter takes a crop as fractions of the frame (x, y, width, height), and the Timeline API places the result in a 1080x1920 canvas with cover, contain, stretch or blur fit. Timeline also joins up to 200 video slots into one MP4. For a prompt-driven edit of an existing clip, Gemini Omni Flash 1.1 accepts a video_url through the Video Router.
None of those steps invents a model. Each runs ffmpeg on a Sume worker or a catalog model you pin by id.
What to tell a client who asked for Ray 3.2
Say plainly that Sume does not run it, then offer the Sume path: pick a catalog id, anchor the first and last frame, chain for length, assemble in Timeline, and deliver MP4. If the brief needs HDR or EXR delivery, Sume's docs do not describe either, so treat that part as out of scope here.
Check the catalog before you plan
Catalogs change, so do not rely on this page for the live list. GET /v1/videos/models returns every video model with its supported_durations, supported_resolutions, supported_aspect_ratios, supported_frame_images, supported_input_references and generate_audio. If a name you want is not in that list, Sume does not run it. Ray 3.2 is not in it.
Pricing follows one rule across the catalog: Sume bills the provider list times 1.25, reserves it at submit, and reports the billable amount as usage.cost on the poll response. The docs do not publish a flat price for each model on the page, so read pricing_skus per model instead of quoting a number from memory.
Sources
Related posts
More in Comparisons
- Ray 3.2 takes 16 keyframes; Sume frame_images takes two
Luma lists up to 16 keyframes per Ray 3.2 clip. Sume's frame_images array uses first_frame and last_frame. Which catalog models list them, per the docs.
- Ray 3.2 tracks 8 faces; Sume swaps 1-4 people with H3 Max Recast
Luma Ray 3.2 lists facial tracking for up to 8 faces. Sume has no tracking output, but h3-max-recast swaps 1-4 people in a clip and Kling drives a still.
- Reference limits per request, vendor vs Sume: Wan 3.0, Seedance, Omni
What each vendor says one video request can take, next to what Sume's catalog allows: Wan 3.0, Seedance 2.5, Gemini Omni 1.1 Flash, and Ray3.2.
- Replicate's default webhook secret versus Sume's signing secret
Replicate's secret is at /v1/webhooks/default/secret and it signs id.timestamp.body. Sume's is at /v1/webhooks/signing-secret and signs timestamp.raw_body.
Written by Sume