Zongze Wu

I'm a research scientist/engineer at Adobe Research in San Francisco.

At Adobe, I work with the FireFly team on Structure Reference.

I got my PhD at Hebrew University of Jerusalem in 2022, under the supervision of Prof. Dani Lischinski and Eli Shechtman from Adobe Research. I got my bachelor degree at Tongji University in 2016.

Email  /  CV  /  Scholar  /  Twitter

My research focuses on generative modeling, spanning diffusion models, GANs, and autoregressive models. I am particularly interested in training efficiency, simpler training frameworks, and efficient architecture design — recent work includes one-step generation via mean flows, representation alignment for diffusion, and separable causal inference in video diffusion.

We show that causal reasoning in video diffusion models is separable from denoising, introducing Separable Causal Diffusion (SCD) that significantly improves throughput and per-frame latency.

Spatial structure, rather than global semantic performance, drives the generation quality of representation alignment in diffusion models.

Improved MeanFlow (iMF) achieves 1.72 FID with a single function evaluation on ImageNet 256x256, closing the gap with multi-step methods using no distillation.

A training-free, inference-time method that uses VLMs to adaptively steer diffusion away from conventional concepts, encouraging novel and creative image generation.

A fast image inpainting method that achieves high-quality results in just 4 diffusion steps, outperforming multi-step methods that require over 50 steps.

Automatically decomposes a diffusion model's visual capabilities into controllable, interpretable directions using low-rank adaptors for concept decomposition and style exploration.

Users can upload an image, and edit the image with natural language. Each edit only takes half a second.

Instead of generating the entire image, we only generate the mask region to facilitate fast inpaint task.

We show StyleGAN3 can be trained with unaligned images, and its w/w+ spaces are more entangled than StyleGAN2.

The child model's latent spaces are semantically aligned with its parent's, inheriting incredibly rich semantics.

Text-based image editing through mapping CLIP space to StyleGAN latent space.

The space of channel-wise style parameters is significantly more disentangled than the other intermediate latent spaces in StyleGAN.

Retrieve foreground images that are semantically compatible with the background.

webpage template from Jon Barron