Evaluation Report: diff_iter1_baseline_256d8

DIFFUSION ITERATION 1: Baseline DiT on 700-canvas dataset ========================================================== Model: ConditionalDiffusionViT (DiT with adaLN-Zero) - embed_dim=256, depth=8, num_heads=8, patch_size=16 - prediction_type=epsilon, beta_schedule=cosine, 1000 timesteps - 200 epochs, batch_size=4, lr=1e-4, weight_decay=0.05 - Data normalized to [-1, 1] - DDIM inference with 50 steps Dataset: 700 canvases (single-action-shoulder-pan-700-combined) - 620 train / 80 val - Canvas: 464x480 (448 visual + 16 motor strip, 6 joints) Purpose: Establish diffusion baseline on same dataset as GPT experiments. GPT best results for comparison: - iter2 (384d12): SSIM=0.756, val_mse=0.0102, action_disc=0.031 - iter6 (384d12+cosine): motor_pos_mae=0.463, motor_dir_acc=85% Diffusion models have key advantages over GPT: - No autoregressive error compounding (all patches predicted in parallel) - Iterative refinement during inference (50 denoising steps) - Stochastic generation (can model multi-modal futures) Key challenges: - Diffusion models need more training epochs - Pixel-space diffusion can be inefficient (no latent encoding) - Quality depends heavily on noise schedule and inference steps