AI Paper Atlas

Research map / Probabilistic and scientific modeling / Variational and generative sampling

Purrception: Variational Flow Matching for Vector-Quantized Image Generation

Răzvan-Andrei Matișan, Vincent Tao Hu, Grigory Bartosh, Björn Ommer, Cees G. M. Snoek, Max Welling, Jan-Willem van de Meent, Mohammad Mahdi Derakhshani, Floor Eijkelboom

ICLR 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Flow matching and flow sampling · 4 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use the vq-ds8-c2i tokenizer instead of vq-f8 (ablation)Purrception DiT-XL/2 · ImageNet-1k 256x256, FID-50k tau=0.8FID: 10.85 -> 6.46Improved
use the vq-ds8-c2i tokenizer instead of vq-f8 (ablation)Purrception DiT-XL/2 · ImageNet-1k 256x256, FID-50k tau=1.0FID: 12.33 -> 9.60Improved
use the vq-ds8-c2i tokenizer instead of vq-f8 (ablation)Purrception DiT-XL/2 · ImageNet-1k 256x256, FID-50k tau=0.7FID: 11.80 -> 7.44Improved
lower inference softmax temperature from 1.0 to 0.8 (tokenizer vq-ds8-c2i) (ablation)Purrception DiT-XL/2 · ImageNet-1k 256x256, FID-50kFID: 9.60 -> 6.46Improved
lower inference softmax temperature from 1.0 to 0.9 (tokenizer vq-f8) (ablation)Purrception DiT-XL/2 · ImageNet-1k 256x256, FID-50kFID: 12.33 -> 10.82Improved
add z-loss regularization (1e-5 log^2 Z) to the training loss (ablation)Purrception DiT-XL/2 · training stabilitytraining loss divergenceImproved
use Purrception instead of CFM and CFM-endpointDiT-L/2 VQ flow · ImageNet-1k FID-10k convergencetraining speed-up: 1.65x fasterImproved
use Purrception instead of DFMDiT-L/2 VQ flow · ImageNet-1k FID-10k convergencetraining speed-up: 3.0x fasterImproved
use Purrception instead of CFM baselinesDiT-XL/2 VQ flow · ImageNet-1k FID-10k convergencetraining speed-up: 2.3x fasterImproved
use Purrception instead of DFMDiT-XL/2 VQ flow · ImageNet-1k FID-10k convergencetraining speed-up: 3.5x fasterImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy