feat: Add Condition Embedding Perturbation (CEP) noise support

[gesen2egee]

Description

This PR implements Condition Embedding Perturbation (CEP) Noise, which is based on the paper:
"Slight Corruption in Pre-training Data Makes Better Diffusion Models" (arXiv:2405.20494).

Adding slight perturbation/noise to condition/text embeddings during training has been shown to improve generative quality, diversity, and prompt adherence.

This feature is integrated into all main training scripts, including:

• train_network.py (LoRA/network trainers for SD 1.5, SDXL, SD3, Flux, Lumina, Anima, etc.)
• Full fine-tuning scripts: flux_train.py, flux_train_control_net.py, sdxl_train.py, sdxl_train_control_net.py, sd3_train.py, lumina_train.py
• DreamBooth (train_db.py)
• Textual Inversion (train_textual_inversion.py, sdxl_train_textual_inversion.py)

The implementation includes filtering logic to ensure that the perturbation noise is only applied to the actual floating-point text/conditioning embeddings, while keeping auxiliary variables like Attention Masks and Coordinate IDs intact.

New Arguments

• --cep_noise (float, default: 0.0): The perturbation noise strength ($\gamma$). Set to 0.0 to disable (default behavior). The paper recommends 1.0.
• --cep_noise_type (str, default: "gaussian"): The type of perturbation noise. Choose between gaussian and uniform.

Application on Small Datasets / LoRA / DreamBooth

CEP Noise serves as an effective regularizer to prevent overfitting, which is a common issue when training on small datasets (e.g., character/style LoRA, DreamBooth).

• Preventing Overfitting: Adding slight perturbations helps prevent the model from memorizing the limited prompts, enhancing the generative diversity of the character/style.
• Tuning for Small Datasets: While the paper recommends $\gamma = 1.0$ for pre-training, a smaller value (e.g., 0.2 to 0.5) is recommended when starting on small datasets to avoid breaking the text-image alignment. If you notice overfitting, you can gradually increase --cep_noise up to 1.0.