How to shoot 1 to 4 person photos for H3 Max Recast
Photo guidance for Recast references on Sume: one photo per new person, framing, light and angle. What the route takes and what to check before you submit.

For Recast, shoot one clear, well lit photo of each new person, facing the camera, with the head and shoulders in frame and the face free of hair, hands and glasses glare. H3 Max Recast takes 1 to 4 reference photos, one for each person you want in the video, and replaces the primary people in the source with them. The photos are the only description of the new people the model gets, so the quality of the photos limits the quality of the swap.
The route documentation fixes the count and the source limits. The advice about framing and light on this page is general photographic practice and not a published requirement of the model, so test it with your own material.
What the route requires
These values are from the Sume catalog.
| Input | Rule |
|---|---|
| reference_image_urls | 1 to 4, one photo for each new person |
| video_url | Source clip, 5 to 30 s, no shot longer than 15 s |
| prompt | Optional, up to 2000 characters |
| Photo address | Public HTTPS URL |
| Resolution | 768p default, or 1080p |
A photo checklist
Shoot a few frames for each person and choose the best, instead of using the first. Small things such as a slight head tilt, the direction of the light and the expression change how a face reads, and ten seconds of choosing saves a failed render. If you can, shoot all the people in the same place with the same light, so the set of references is consistent.
Treat each photo as an ID picture for the model. The more clearly it shows the face, the less the model has to guess.
- One person per photo, so there is no doubt who the photo stands for.
- Face forward or a slight turn, with both eyes visible and the whole face in view.
- Soft, even light with no hard shadow across the face and no strong backlight.
- Neutral or mild expression, because an extreme expression can appear in every frame.
- Sharp focus and a large enough file, not a screenshot of a screenshot.
- No sunglasses, mask or heavy filter unless the person must wear them in the video.
Match the photo to the source shot
Look at the person in the source before you shoot the replacement. If the source shows a person in profile, in low light and far from the camera, a bright frontal photo is still the best reference, but the result will show that the source is limited. Compare the apparent age, the hair and the build in the photo with those in the source, since the swap has to fit the body and the motion that remain. If you swap more than one person, order and label your photos in the same order as the people appear, and make a note of which photo stands for which person.
See swapping two people in one clip for the case with more than one person.
Consent and a test run
Keep the original, uncropped files. If a result needs a different crop, you will want the full frame and not a crop of a crop, and a lossless original gives the model the most detail to work with.
Only use photos of people who agreed to appear, and keep that agreement with the project. Before you run the full ad, cut a short, five to eight second piece of the source where the person is clearly in frame, and run Recast on it at 768p. A 5 second source at 768p is $1.875 before rounding up to cents (5 seconds at $0.375). If the face holds up there, run the full length. If it does not, change the photo first and not the prompt, because the photo is what describes the person.
The path from a tested photo to a market version is in one ad, two Recast jobs.
Sources
Related posts
More in Use cases
- HVAC seasonal promo: 12 neighborhood versions in one Format bulk run
A heating and cooling company can queue a furnace-tune-up promo for 12 neighborhoods in one bulk run: concurrency, per-item input, and a spend cap per run.
- Image-to-image vs image edit on an API: one route, five fields
On Sume, image-to-image and image edit are one POST /v1/images call. The fields that decide the result: input_references, aspect_ratio, mask_url, quality, n.
- Indie skincare brand launch kit: texture still, serum-drip still, reel
An indie skincare brand can make a texture still, a serum-drip still and a studio reel from one packshot with three Sume catalog Formats. What each one does.
- Instagram for TV timeline: Fire TV test to Google TV, what to do
Reels reached the living-room screen in a US test on Fire TV, then Google TV. The dated facts, what is unknown, and a short checklist for Reel makers.
Written by Sume