Results from an experiment using Condition Embedding Perturbation. https://arxiv.org/pdf/2405.20494
Code is very simple, add noise during the training loop between the text_encoder forward and the Unet prediction based on this code:
perturbation_deviation = embedding_perturbation / math.sqrt(encoder_hidden_states.shape[2])
perturbation_delta = torch.randn_like(encoder_hidden_states) * (perturbation_deviation)
encoder_hidden_states = encoder_hidden_states + perturbation_delta
"embedding_perturbation" is the tunable gamma from the paper.
Models are SD1.5 trained for 30 epochs, Unet only, AdamW8bit, 2e-6 constant LR, batch size 12. EveryDream2Trainer config json included.
#!bin/bash
# python train.py --config train_ff7_emb_pert000.json --project_name "ff7r embedding_perturbation 0.0" --embedding_perturbation 0.0
python train.py --config train_ff7_emb_pert000.json --project_name "ff7r embedding_perturbation 1.0" --embedding_perturbation 1.0
python train.py --config train_ff7_emb_pert000.json --project_name "ff7r embedding_perturbation 1.7" --embedding_perturbation 1.7
python train.py --config train_ff7_emb_pert000.json --project_name "ff7r embedding_perturbation 3.5" --embedding_perturbation 3.5
