Twitxy Saved

Pedro

Verified account This account is verified. Learn more

@pmpcurvo

PhDing @UvA_Amsterdam | Generative & World Models
Amsterdam pedrocurvo.com Joined May 2025
694 Following 283 Followers 281 Posts

Latest posts 5

Pinned post

Pedro Verified account @pmpcurvo
鈾撅笍 LiFT: Loop Flow Transformers 馃У馃憞

Diffusion models usually spend more inference compute by taking more sampling steps. What if we could also spend compute inside each step?

LiFT turns part of a DiT into a recurrent core, letting us scale depth at inference with only a small architectural change. On ImageNet, it beats dense DiT-XL/2 with 3.34 lower FID, while using 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.

Jointly led with Mohammad Mahdi Derakhshani 馃憫(@mmderakhshani)

And amazing supervisors Cees Snoek (@cgmsnoek), Jan-Willem van de Meent (@jwvdm) and Gertjan J. Burghouts (@gjburghouts)
1:29
9 replies 52 reposts 360 likes 24K views View on X
Pedro Verified account @pmpcurvo
a must-read for anyone interested in continuous diffusion and flow-based language models
Floor Eijkelboom Verified account @FEijkelboom Nov 9
Our survey on continuous diffusion and flow language models is out! 鉂わ笍

A little animated tour of what鈥檚 inside 馃憞 From representations and training to few-step generation, scaling, and open problems.

continuous-lm.github.io
View quoted post on X
Pedro Verified account @pmpcurvo
Imagine scaling a generative model at inference simply by running it deeper, instead of always making it larger. With LiFT, we show that looping actually pays off: our L/2 model reaches a 3.3 lower FID than our dense DiT-XL/2 baseline, while using 60% fewer parameters and 52% fewer inference FLOPs.

Very happy to have worked on this with Mohammad!
Mohammad Mahdi (Medi) Derakhshani Verified account @mmderakhshani Oct 6
LiFT: Loop Flow Transformer: A correspondence between flow matching and looped DiTs.

Finally cracked it with Pedro. I do think it is fully compatible with multimodal setups too. We'll release the codebase, W&B artifacts, and checkpoints. Huge shout-out to the team!
View quoted post on X
0 replies 0 reposts 9 likes 1.1K views View on X
Pedro Verified account @pmpcurvo
Guide with examples, not rewards 馃悩

Controlling what a pretrained generative model produces is still mostly a choice between three slow options: fine-tune it, attach a reward network, or search at inference. We found flow matching allows a fourth, and it costs almost nothing.

In deterministic interpolants, the velocity of the flow is determined by where the trajectory is headed: the endpoint mean. Shift that mean, and the entire flow shifts with it.

This turns control into a matter of reference. Change the examples that define the endpoint, and you change the direction the model follows. The examples need not be perfect. They only need to point the flow toward the attribute you want.
Color, identity, style, and structure, all controllable through examples. 馃У馃憞