Spokester1 ViT-B

• 2x16x16 patches
• trained for 267,000 steps
• 120k steps at 224x224
• rest at 320x320
• variable temporal window size, from 2 frames to 32 frames
• NoPE, RoPE2d with RoPE box jitter
• conditioned on noisy and downscaled-upscaled depth and optical flow
• No CFG
• velocity loss, with model outputing a prediction of the clean input