Residual-Stream Burden Shapes Representation Learning in Diffusion Transformers
In diffusion-based generation, a neural network can be trained to predict the clean data, the noise, or the velocity from a noisy input, yet plain Diffusion Transformers operating on large pixel patches succeed with clean prediction and fail with noise or velocity prediction.
Tong-Tong Liang, Si-Qi Kou, Zi-Qiao Xi et al.
· 0 citations