Evaluation Report: diff_iter3_sample_fixed_inference

DIFFUSION ITERATION 3: Sample prediction + higher LR + 500 epochs ================================================================== Changes from Iteration 2 (still producing noise): - prediction_type: epsilon -> sample (predict clean image directly) - epochs: 300 -> 500 (diffusion needs much more training) - lr: 3e-4 -> 5e-4 (higher peak since sample prediction is easier to optimize) - warmup_epochs: 10 -> 20 (longer warmup for stability at higher LR) Rationale: Iteration 2 reduced denoising loss from 0.675 to 0.506, but inference still produces noise (SSIM=0.008). The issue: epsilon prediction asks the model to predict the noise component, which is harder to learn than predicting the clean image directly. With "sample" prediction: - The model predicts x_0 (clean patches) instead of epsilon (noise) - More intuitive: model learns "what should this look like?" not "what noise was added?" - Converges faster in practice for conditional generation - The DDIM step still works: uses pred_x0 directly instead of deriving it from epsilon Also, this canvas has an asymmetry: context patches are ALWAYS clean while last-frame patches are noisy. Epsilon prediction struggles with this because the model must learn different behaviors for context vs. target regions. Sample prediction is more uniform: predict clean for everything, training only optimizes the last-frame predictions. The previous training ran for 300 epochs but loss was still decreasing. 500 epochs should allow better convergence.