huber-exp

SD1.5 experiments with Huber and MSE loss. All models trained for 4 epochs on approximately 250k images from a variety of sources. Approximately half from LAION Aesthetics, and a few thousand 4K video rips with COG-VLM captions.

Trained using Everydream2 Trainer (https://github.com/victorchall/EveryDream2trainer) on an RTX 6000 Ada 48gb. Each epoch takes approximately 10 hours for a total of about 40 hours per model.

• Multi-aspect ratio trained with nominal size of <=768^2 pixels for each bucket
• Batch size 12 with grad accum 10.
• AdamW 8bit optimizer with standard betas of (0.9,0.999) and weight decay of 0.010.
• Automatic mixed precision FP16 (note: grad scalar val was surprisingly identical on all runs)
• TF32 matmul and SDP Attention
• 3.0e-6 LR cosine schedule with a ~12 epoch target to decay, ending around 2.3e-6 at end of training
• Pyramid noise using discount 0.03
• Zero offset noise of 0.02
• Min SNR gamma of 5.0
• Unet only training, text encoder left frozen.
• Conditional dropout of 10%

The following models were produced:

• 768_huber.safetensors - Huber loss only
• 768_mse_plus_huberd1.5.safetensors - MSE Plus Huber (d=1.5) loss
• 768_ts0huber_ts999mse.safetensors - Huber loss at timestep 0 interpolated to MSE loss at timestep 999
• 768_ts0mse_ts999huber.safetensors - MSE loss at timestep 0 interpolated to Huber loss at timestep 999

Worth noting timestep 0 is the lowest-noise-added step and 999 is most noised timestep