A Scene-Relighting Workflow

When you relight a scene by submitting a new prompt with different lighting keywords, the generative model produces a new image that shares the broad concept of the original but is not structurally identical. The subject's face will have slightly different proportions, the background objects will move, and texture details will be inconsistent. For product photography this is especially damaging — the product's surface detail, label text, and dimensional proportions must be identical between lighting variants because the images will appear side by side on a product page or in a comparative ad.

Depth-conditioned relighting avoids this problem by treating the lighting change as a localized edit rather than a new generation. The depth map extracted from the original image encodes the three-dimensional structure of the scene — how far each pixel is from the camera. The relighting model uses this depth map as a constraint, ensuring that surfaces facing the new light source brighten in proportion to their angle toward that source (Lambertian shading) and that occluded surfaces receive correct shadow while maintaining the underlying geometry. The result is a relighted image where every surface is exactly where it was in the original, only the illumination has changed.

The scene-relighting workflow consists of three sequential stages. Stage one runs two parallel preprocessing nodes: a Depth Estimation node that converts the input photograph into a grayscale depth map, and a Surface Normal Estimation node that converts the same photograph into a normal map encoding the facing direction of every surface in the scene. Both nodes feed into stage two.

Stage two is the Relighting Model node. It receives three inputs: the original image, the depth map, and the normal map. It also receives a light specification: light direction as azimuth and elevation angles (for example, 210° azimuth and 35° elevation for a dramatic side-back light), light color temperature in Kelvin (3200K for warm tungsten, 6500K for cool daylight, 8000K for overcast blue), and light intensity as a multiplier relative to the ambient level detected in the original image. The node applies these parameters to synthesize new illumination that is geometrically consistent with the reconstructed scene structure.

Stage three is a Blend and Refinement node that composites the relighted output over the original using a structure-preserving blend — high-frequency texture detail from the original image is layered back over the relighted output to recover fine surface texture that may be smoothed by the relighting model. This step preserves label text, fabric weave, and skin pore detail that would otherwise be lost.

The light specification parameters map directly to real-world lighting scenarios. For product photography, the standard "beauty light" position is approximately 30° azimuth and 45° elevation with a color temperature of 5500K and an intensity multiplier of 1.2 — slightly brighter than the original ambient to make the product pop. For dramatic editorial portraits, a side light at 90° azimuth and 20° elevation with a color temperature of 4200K and intensity 1.4 creates a Rembrandt-style effect. For e-commerce lifestyle shots that need a "golden hour" feel, set azimuth to 260° (setting sun direction), elevation to 8°, color temperature to 3000K, and intensity to 0.9.

The relighting model in Floniks also accepts a secondary light specification — a fill light, typically from the opposite side of the key light at lower intensity. Configure the fill light at the key light's opposite azimuth, elevation 10–20° lower, intensity 0.3–0.4 of the key, and color temperature 500–1000K cooler. This two-light setup (key + fill) prevents the dramatic shadow side of the subject from going completely black, which is the tell that distinguishes AI relighting from professional photography lighting.

The most visually complex aspect of relighting is managing specular highlights — the bright point or area of reflection that appears on shiny surfaces (glass, metal, glossy product packaging, skin highlight) when a light source is at the correct angle. The depth and normal map approach handles diffuse (matte) surface relighting well but often over-brightens or misplaces specular highlights because it does not model the specular reflectance properties of each material.

To address this, add a Specular Mask node after the depth and normal estimation stage. This node analyzes the original image to identify specular-highlight regions — areas where luminance is significantly above the average surface brightness. It outputs a binary mask that tells the relighting model "these pixels are specular-dominant; apply lighter relighting treatment here." Configure the specular threshold at 0.80 of the maximum luminance in the image. For highly reflective product photography (glass bottles, chrome hardware), also enable the glossy surface mode in the relighting node, which treats these regions as mirror-like reflectors that need new environment reflections generated at the new light position rather than simple luminance scaling.

Many commercial use cases require delivering the same image under multiple lighting conditions: a product shot in studio lighting, in warm storefront lighting, in cool outdoor daylight, and with a dramatic dark-studio spotlight treatment for premium positioning. Rather than running the relighting workflow once and manually re-configuring light parameters four times, build a batch variant node in Floniks that holds an array of light specification presets and fans out to four parallel relighting branches, each receiving the same depth and normal map inputs but a different light configuration.

This batch topology means the preprocessing stage (depth and normal estimation) runs only once, and the computationally lighter relighting pass runs four times in parallel. The four outputs route to a Comparison Composite node that arranges all variants in a 2×2 grid for client review. Once the preferred lighting is approved, the workflow stores the approved light specification as a named preset that can be applied to the full catalog of product images without further review — standardizing the lighting treatment across hundreds of SKUs in a single workflow run.