Anand Bhattad (@anand_bhattad)
2024-12-05 | โค๏ธ 201 | ๐ 33
[1/7] Iโm super stoked to share LumiNet! ๐
๐กLumiNet combines our latent intrinsic image representations with a powerful generative image model. It lets us relight complex indoor images by transferring lighting from one image to another!
Most importantly, it is completely data-driven - no labels, no marked-up data, not even depth or normals. No physical model either! Just two images in, one relit image out! โจ
Some of the results are astonishing! ๐คฏ Check out how it handles complex lighting - inter-reflections on TV screens, cast shadows, soft shadows - everything! The model seems to โknowโ where light sources are and how light bounces off different surfaces. This canโt just be mere corraltions and pattern matching - itโs doing some serious spatial reasoning! ๐ฏ
๋ฏธ๋์ด
์์ฝ
LumiNet์ ๋ ์ด๋ฏธ์ง๋ง์ผ๋ก ๋ณต์กํ ์ค๋ด ์กฐ๋ช ์ ๋ค๋ฅธ ์ฅ๋ฉด์ ์ ์ดํด ์ฌ์กฐ๋ช ํ๋ ๋ฐ์ดํฐ ๊ธฐ๋ฐ ๋ชจ๋ธ์ ๋๋ค. ๋ผ๋ฒจยท๊น์ดยท๋ ธ๋ฉยท๋ฌผ๋ฆฌ ๋ชจ๋ธ ์์ด๋ ๋ฐ์ฌ๊ด, ์บ์คํธ/์ํํธ ๊ทธ๋ฆผ์ ๊ฐ์ ์กฐ๋ช ํจ๊ณผ๋ฅผ ์์ฐ์ค๋ฝ๊ฒ ์ฌํํฉ๋๋ค.
๐ Related
Auto-generated - needs manual review