LightShift

LightShift: Controllable Indoor Relighting via Video Diffusion Priors

This repository contains the fine-tuned UNet checkpoint for LightShift, a controllable indoor relighting method that repurposes Stable Video Diffusion for image-based relighting.

Usage

This checkpoint is compatible with the Diffusers library. See the code and inference scripts at Pensioner-11/LightShift.

from diffusers import StableVideoDiffusionPipeline

pipe = StableVideoDiffusionPipeline.from_pretrained(
    "Pensioner/LightShift",
    torch_dtype=torch.float16,
).to("cuda")

Model Description

• Base model: stabilityai/stable-video-diffusion-img2vid-xt
• Fine-tuned component: UNet
• Training steps: 200,000

Citation

Citation information will be updated after publication.