Ray 3.2 tracks 8 faces; Sume swaps 1-4 people with H3 Max Recast

Luma lists facial performance tracking for up to 8 faces in Ray 3.2. Sume does not run Ray 3.2 and returns no face-tracking data, but it has two face-related tools: H3 Max Recast replaces 1 to 4 people in a source clip with people from photos, and Kling 3.0 Motion Control drives a still image with the motion of a reference video.

Those do different jobs. Tracking is about following performances in a generation; recast and motion control are about changing who appears in a shot.

What Luma lists

The Ray 3.2 launch page (read 2026-10-05) lists facial performance tracking, up to 8 faces, alongside 16 keyframes, 1080p and clips up to 20 seconds. The page is the only source used here for Ray 3.2 claims.

Recast: replace the people in a clip

In the Video Router docs, h3-max-recast replaces the people in a source video_url with the people from 1 to 4 reference_image_urls, with one photo per person. It runs at 768p or 1080p, accepts 5 to 30 seconds, and takes an optional prompt. The duration is the source length.

If your source has two people, send two photos. The model replaces each person, which is closer to a casting swap than to tracking.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: recast-001" \
  -d '{"model":"h3-max-recast",
       "video_url":"https://example.com/source.mp4",
       "reference_image_urls":["https://example.com/person-a.jpg","https://example.com/person-b.jpg"],
       "resolution":"768p","mode":"async"}'

Motion control: a still performs a reference

Kling 3.0 Motion Control (POST /v1/kling/3.0/motion-control, public id kling/3.0/motion-control) takes one still (image_url, or a ready avatar) and a motion_video_url. The reference video drives all motion in the output. The hosted MCP tool is kling-motion-control_create, listed in the MCP tools and gates docs.

This is the nearest Sume answer to a performance transfer, though it is not facial analysis: you get a video, not tracking data.

Side by side

Choosing between the two Sume paths

Use recast when you already have footage and want different people in it. Use Kling motion control when you have a character still and a performance video, and want the still to move like the performance. If you need more than 4 people, split the shot into separate jobs and join them in Timeline.

Neither tool promises identity preservation beyond what the model does with your photos, so test on a short clip before you commit a long one. Each job reserves its price at admit and uses the Sume list times 1.25 rule, so a short test is cheap.

A cost check before a recast run

H3 Max Recast is billed at the provider list times 1.25 like each other catalog model, and its duration is the source length, so a 12-second source means 12 billed seconds. Trim the source to the shot you need first with video trim to avoid paying for seconds you will cut anyway.

Recast accepts 5 to 30 seconds. A source shorter than 5 seconds is outside the range, so pad or choose a longer range.

What to verify in your own test

• Does each replaced person keep the right position in the frame across the whole clip?
• Does a person who leaves and re-enters the frame keep the same photo identity?
• Does the clip's audio survive, or do you re-add it in Timeline?
• Do hands and profile angles hold up at 768p versus 1080p?

What this page does not claim

It does not claim that Sume tracks faces, that recast matches Ray 3.2's output, or that either tool preserves an identity to a given standard. It reports what the docs say each Sume surface accepts and what Luma's own page lists. Run a short test with your own footage and photos before you commit to a campaign.