E-TSDM: Early Timestep-shared Diffusion Model

[Koratahiu]

This draft implements the Early Timestep-shared Diffusion Model (E-TSDM) strategy proposed in the paper:
Lipschitz Singularities in Diffusion Models

This method aims to alleviate "Lipschitz singularities" (infinite rates of change) that occur near the zero timestep, resulting in improved training stability and generation quality. The paper observes that the network's Lipschitz constant with respect to time explodes near $timestep=0$ (e.g. timesteps 0 to 100), causing numerical instability.

E-TSDM addresses this by sharing timestep conditions within small sub-intervals in this region (mapping the model input to the start of the interval while preserving the physical noise schedule), effectively forcing the Lipschitz constant to zero within those steps.

I think this issue applies to all prediction targets (eps, v and flow matching), and this method works for all of them.

Includes: #1224 #1225

TODO

• To be tested
• Add UI (for now it's enabled by default)
• Add for all models, for now only: SD, SDXL, Chroma, and Flux are supported (minor change for each, anyway)