Expand description
The LTX-2.5 spatial latent upscaler — the ×2 between the two denoising stages.
Structurally simple and computationally not: 3-D convolutions at 1024 channels with ordinary zero padding, GroupNorm(32) and SiLU, four residual blocks before a pixel-shuffle upsample and four after. It runs on the un-normalized latent — the video VAE’s per-channel statistics are undone going in and reapplied coming out — because the upscaler was trained in the VAE’s own units, not the diffusion model’s.