Foveated Diffusion: Efficient Spatially Aware Image and Video Generation

Image generation demo for Foveated Diffusion. The base model is FLUX.2-klein-base-4B.

Project website · Model weights · Paper

LoRA model

Finetuned with random gaze locations during training. Most general-purpose foveation LoRA.

Active: random

Foveal region (circular) — place a circular HR region on the image. cx, cy ∈ [-0.5, 0.5] (0 = image center); r is relative to half the image diagonal.

-0.5 0.5
-0.5 0.5
0 1

Tokenization mask — white = HR (foveal) tokens, gray = LR (peripheral). Refreshes as you move the circle sliders. Generated image — appears once denoising finishes.