Rewire the light without rewiring the outlet

An IC-LoRA for LTX-2.3-22B redirects the sun in a scene using an embedded gray sphere. A geometric method for post-production.

Changing the time of day on a shot that has already been filmed is what a new IC-LoRA trained on LTX-2.3-22B offers. Its originality lies less in the result than in how it is given its commands. To indicate where the sun should come from, there are no parameters to fill in. A small gray sphere is overlaid in the upper right corner of each image, lit from the desired direction and casting its shadow on the opposite side. This ball is the control signal: the model reads it and relights the scene so that the main light aligns with it.

A short caption completes the instruction, using a fixed template made of a trigger, a rendering, and a direction. Twelve renderings were trained, from low-angle sunlight to the diffused light of an overcast sky, including backlighting, and eight directions divide the area around the subject into forty-five-degree increments. The elevation of the sphere also sets the height of the sun—low, medium, or high.

Processing is done in a single pass at the target resolution, without a separate relighting model or frame-by-frame calculation. A ready-to-use ComfyUI workflow accompanies the release and handles both the rendering of the sphere and its overlay, via a node created by Eric Venti. Outside of ComfyUI, the operation becomes manual again, with the publisher detailing the expected dimensions and the corresponding command line.

The limitations are stated straightforwardly. Training was conducted only on exteriors; interiors are out of scope. The tool relights, it does not restyle. No masks are supported. And the most common error relates to the sphere itself: if it is missing, cropped by a late resize, placed in another corner, or absent from certain frames, it no longer controls anything, and the relighting becomes weak or inconsistent. When the caption contradicts the ball, the ball wins.

Training was performed at rank 32 over three thousand steps, using a proprietary dataset of before-and-after pairs, with the LTX-2 Community Trainer. Released on Hugging Face in safetensors format under the LTX-2 community license, with access conditioned on sharing an email address and a username.