FlashRelight

Official inference checkpoints for FlashRelight: Portrait Video Relighting with Dynamic Lighting (NeurIPS 2026). Paper · Code

FlashDelight suppresses the original lighting in a portrait video. FlashRelight uses the delighted clip to synthesize lighting from an HDR environment map or a reference video. Text control works best with lighting descriptions similar to the training captions. The released models process 49 frames at 480 × 832 pixels.

Checkpoint	Size	Use
flashdelight_rank0_full.pt	6.93 GB	First stage. Use it on its own for delighting or as the first step in relighting.
flashrelight_rank0_full.pt	7.04 GB	Second stage. Use it with FlashDelight for relighting.

Download

From the FlashRelight code repository, download both checkpoints to the weights directory.

hf download JiahuiSheng/FlashRelight config.json flashdelight_rank0_full.pt flashrelight_rank0_full.pt --local-dir weights

For standalone delighting, download the configuration and first-stage checkpoint.

hf download JiahuiSheng/FlashRelight config.json flashdelight_rank0_full.pt --local-dir weights

Both stages use the shared backbone settings in config.json. The inference code applies the stage-specific input dimensions and LoRA settings. Both modes also require the Wan2.1 VAE, text encoder, and tokenizer. Wan2.1 base files

SHA-256 checksums

04925b6ca8a4b537685a2791be65d0a51d2025fba8d2cb4ae17d4bf9530c1f5c  flashdelight_rank0_full.pt
8b9989cdebd2d9815f3678baf4493fbf7c0ae65fe2c74756769e4f4df1c226d7  flashrelight_rank0_full.pt