ReaDiT Guidance: Control for Image and Video Generation
using Diffusion Transformer Features

Jay Maanav Mahajan1*,  Chang Liu1*,  Rauf Makharov1,  Viraj Shah2,  Alex Schwing1†,  Svetlana Lazebnik1†
1UIUC   2Google
*Equal contribution    †Equal advising
arXiv (coming soon) Code (coming soon)

Abstract

We present DiT Readout (ReaDiT) Guidance, a lightweight framework for controlling generation with Diffusion Transformer (DiT) models via their internal feature representations. ReaDiT Guidance uses features from a single DiT block to steer the generative process according to spatial targets — depth, pose, or edge maps — provided at test time. Since modern text-to-video models are largely built on DiT backbones, ReaDiT Guidance naturally extends to video generation, enabling camera and motion control. Our approach achieves competitive or improved results compared to existing feature-based and adapter-based approaches while requiring significantly fewer parameters.

Approach

Method overview — figures/Method_figure_draft2.png
Method Overview. Left: ReaDiT takes a single intermediate DiT feature and guides generation by optimizing the latent zt. Right: Training — noise an image to timestep t and supervise with an off-the-shelf target.
ReaDiT architecture — figures/Readout_arch3.png
ReaDiT Architecture. Input DiT features are encoded into multi-resolution maps and progressively fused to produce spatial outputs.

Spatial Prediction

ReaDiT predicts dense spatial maps from frozen DiT features at any denoising timestep.

Spatial Image Guidance

Given a spatial guidance map, ReaDiT steers generation to match the target layout while maintaining image quality and prompt adherence.

Dual Guidance

ReaDiT can jointly apply depth and pose guidance. Even when control maps come from different source images, it automatically decouples the two signals to produce coherent results.

Inconsistent — depth & pose from different source images

Depth reference

Pose reference

Dual guidance result