H3 Max Relight: What the Lighting Sphere Actually Controls

H3 Max Relight changes the lighting of an existing video. It does not generate a new scene from a text prompt. The current fal route takes two required inputs: a source video and an image of a lighting sphere. Its job is to transfer the sphere's illumination while preserving the source subjects, motion, camera, and audio.

That makes Relight useful after the shot already works. If the performance, framing, and timing are wrong, relighting is the wrong repair. If those parts are right but the light feels flat, cold, or inconsistent with the next shot, this is the route to consider.

The lighting sphere is the instruction

The reference image is not a moodboard, a product photo, or a frame from another movie. The API calls it a lighting sphere image: a rendered sphere that describes how light should fall across a three-dimensional surface.

A useful sphere makes four things legible:

• the direction of the main light;
• whether that light is hard or soft;
• the color of the key light and the surrounding fill;
• bright reflections, dark falloff, and any colored rim light.

Think of it as a compact lighting diagram. A warm highlight on the sphere's upper right and a dim blue edge on its left gives the model a spatial instruction. “Cinematic sunset” does not. Relight's published request schema has no text-prompt field, so prose cannot rescue an ambiguous sphere. The fal playground also includes a Lighting Studio for rendering this input instead of asking users to improvise a sphere in an ordinary image editor.

This also suggests a better way to iterate: keep the source clip fixed and change one part of the sphere at a time. Moving the highlight tests direction. Widening it tests softness. Changing its color tests palette. Replacing everything at once makes it harder to identify why a result changed.

What should remain from the source video

fal describes Relight as preserving the source subjects, motion, camera, and audio. Those are the continuity anchors. In practical terms, review the output against the input rather than judging it as a completely new clip.

“Preserve” is the provider's intended behavior, not a promise that every pixel will survive. The safest source is easy to read before the model starts: one clear subject, coherent exposure, limited motion blur, and no crushed shadows hiding important features. Relighting cannot faithfully reveal detail that the source never recorded.

Prepare the end of the clip, not just the beginning

The input contract contains an easy-to-miss timeline rule. Source video can be up to 15 seconds. It is converted to a constant 24 frames per second, then trimmed at the end to the closest frame count matching 17n + 5. At least 56 frames—about 2.33 seconds at 24 fps—must remain. The output follows that prepared frame count and duration.

The practical consequence is simple: do not place the only important gesture, final word, or product reveal in the last few frames. The tail can be shortened during preparation.

Before sending a clip, leave a little visual and audio breathing room after the meaningful action. If exact synchronization matters, inspect the prepared result frame by frame and confirm the final syllable, beat, and pose are still present. A nominal “five-second source” is not a guarantee that every original end frame becomes part of the output.

A source-shot brief that leaves room for relighting

When the source does not exist yet, make it structurally simple. Here is an original, untested prompt for the standard generator on this site:

Five-second locked medium close-up of a ceramic perfume bottle on a dark stone table. The bottle remains still while a thin ribbon behind it moves gently in one continuous shot. Keep the label readable, the background uncluttered, and the exposure neutral with soft, even illumination. No cut, zoom, flicker, moving spotlight, or dramatic color cast. Quiet room tone and a subtle fabric rustle.

The prompt deliberately avoids dramatic lighting. It asks the source generation to solve composition, identity, motion, and sound first. A separate lighting sphere can then carry the stronger visual decision.

This site's generator cannot run the Relight endpoint or accept a source-video upload. It can, however, create and download a clean text-to-video or image-to-video source at 480P or 768P.

Choose resolution after the lighting test

The current minimax/h3-max/relight schema offers 480P, 768P, 1080P, and 2K. The documentation labels 480P and 768P as native options; 1080P and 2K use latent upscaling and refinement from 768P. Higher output therefore costs more, but it does not add source detail that was absent from the input.

fal's published Relight rates, checked October 16, 2026, are:

fal says it bills per second of generated video and gives the five-second examples shown above. A generated 12-second 768P output would therefore cost $0.96. The API separately documents its frame-preparation trim, but it does not spell out fractional billing behavior after that trim; check the actual generated duration when reconciling a job. These are fal API rates, not this site's credit prices.

Start with 480P or 768P while deciding whether the sphere is doing the right job. Move to a refined resolution only after light direction, contrast, and color are correct. Paying four times the 768P rate for 2K does not make an unclear lighting reference clearer.

The schema also exposes adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 aspect-ratio choices. Treat a ratio change as a composition decision, not merely an export setting. If preserving the original frame is important, inspect the result for cropping or newly synthesized edges before accepting it.

A relighting pass needs a narrow acceptance test

Write the pass conditions before submitting the job. For the perfume shot, a useful checklist could be:

1. The label and bottle silhouette remain stable through all frames.
2. The highlight arrives from the side shown on the sphere.
3. The ribbon motion and camera framing match the source.
4. The new shadow does not flicker or jump across the table.
5. The ending pose and final audio beat survive source preparation.
6. The audio track remains aligned and audible.

If the identity or action changes, compare the source rather than editing the sphere. If only the direction or color of light is wrong, revise the sphere. That separation prevents an expensive loop in which the lighting reference is blamed for a weak source—or the source is regenerated because one highlight is misplaced.

Relight is a finishing route, not a universal generator

The route name is minimax/h3-max/relight. It requires video_url and reference_image_url; the latter is specifically the lighting-sphere image. Optional controls include output resolution, aspect ratio, seed, safety checking, and a data-URI response mode. The response contains a video and seed, with timings when available.

There is no prompt input in the current contract. There is also no instruction to swap a person, replace an object, rewrite motion, or extend the timeline. Those are different jobs with different endpoints. Relight earns its place by keeping the shot and changing one production layer: illumination.

That is the cleanest mental model for the whole workflow:

Source video = subject + motion + camera + audio
Lighting sphere = direction + softness + color + contrast
Relight output = the source shot under the new lighting plan

Use the image-to-video workflow when you still need to invent the motion. Use the custom-audio guide when sound is the unresolved layer. Use Relight only when the clip is already worth keeping and the light is the part you want to change.