RelightFormer: a feed-forward generative transformer for multiview object relighting
PolyUHK · hf · 2026-09-09
PolyUHK researchers propose RelightFormer, a feed-forward generative transformer that directly relights single- and multi-view object images.
- Target illumination is injected via cross-attention
- Permutation-invariant encodings handle unordered views without fixed ordering
- Trained on a large synthetic dataset
Compared to optimization-based relighting pipelines, the feed-forward design enables fast inference for both single-image and multiview relighting.
More from Multimodal
- One-Take Rainy Toll Booth: A Full AI Video Prompt for Korean Supernatural Mystery — umesh_ai · 2026-09-09
- MiniMax H3 Ref2V resists surgical video edits, users hunt for prompting workarounds — Weird_Ad4978 · 2026-09-09
- Full AI video pipeline: Qwen 3.8 + Flux 2 Klein + Minimax H3 + LTX 2.5 + Breeze TTS — CQDSN · 2026-09-09
- Minimax H3 videos look h264-compressed regardless of settings, Redditor reports — FoxTrotte · 2026-09-09
- Mage brings MiniMax H3 & H3 Turbo with unlimited video generation, LoRA support — MiniMax_AI · 2026-09-09
- One-line prompt test shows GPT Image 2.5 producing near-indistinguishable iPhone-style photos — Scobleizer · 2026-09-09