Atul (@atulku) profile picture
Geneva, Switzerland
Joined September 2009
1/ What happens when a denoiser trained at one noise level is reused inside an iterative sampler? We trained SwinIR with Noise2Noise at σ=10, then dropped it unchanged into the constrained sampler of @ZKadkhodaie & @EeroSimoncelli (NeurIPS ’21) for 10% random inpainting. Baseline SwinIR: 6.08 dB on Set12. SwinIR-WNE: 23.87 dB. Same backbone. Same N2N pairs. Same sampler. With François Fleuret @francoisfleuret. ICML 2026. 🧵
3
16
1
64
24,039
There are densities for which a simple analytically defined interpolation from a normal results in mods "growing out of nowhere" instead of "sliding" there. This makes the transport field super hard to learn since it cannot just follow them.
Replying to @balintmt
While it works nicely in certain special cases, it fails in others. (7/9)
1
2
1
6
3,065
Preprint on interpolating between unnormalized densities: arxiv.org/abs/2301.07388 We introduce an objective for training continuous NFs without samples but with a given, non-normalized target energy that we can evaluate pointwise. Joint work with @francoisfleuret. (1/9)
3
11
60
13,065
Napping is the best part of every day
105
2,667
1,084
Asking the right questions takes as much skill as giving the right answers. #copied
1
1
22 Unbelievable Places that are Hard to Believe Really Exist :D boredpanda.com/amazing-place…