DAGS: Disentangled Appearance-and-Geometry Steering of a Frozen Image DiT for Temporally Stabilized Generative Rendering

arXiv cs.GR details DAGS, a new conditioning scheme for diffusion transformers that targets one of generative rendering’s biggest pain points: unstable results over time. The method keeps the image DiT backbone frozen and steers it with separate appearance and geometry signals, which is the kind of design that should matter to graphics programmers trying to preserve pretrained priors while adding control.

The core idea is deliberately lightweight. Two small convolutional encoders compute conditioning features once per frame and inject them as learned, per-layer residuals into image tokens, avoiding the quadratic cost of stacking conditions through attention. Because the control path lives outside the backbone, the approach also sidesteps the usual risk of overfitting the base model away from its pretrained behavior.

DAGS also adds a small recurrent lighting stabilizer plus a training-free temporal guidance term, turning a per-frame image model into something closer to a streaming renderer. The authors say it is not real-time, but it aims to trade compute for controllability and quality in a way that could be useful for offline rendering, look-dev, or hybrid pipelines.

On a matched 1-spp + G-buffer input, the system reportedly improves reconstruction by +8.6 dB and +10.1 dB PSNR over Intel OIDN and RGB<->X, while being 2.5-8x more temporally stable by temporal-LPIPS flicker. For teams exploring neural denoising or generative rendering, the practical takeaway is that disentangled control and temporal stabilization may be a better fit than bolting more attention onto...