README.md

T2ITrainer

T2ITrainer is still under development stage and not stable yet.
It updates very frequently, please check the change logs for details.

Prerequisites

• PyTorch version >= 2.3.0 with CUDA 12.1 support (torch>=2.3.0+cu121).

Supported Training Configurations

• For Lora training: Supports kolors, sd3, and hunyuandit 1.1.
  • Hardware requirements:
    • CUDA GPU with >13GB memory for hunyuandit lora training.
    • CUDA GPU with 11GB memory for kolors training.
    • CUDA GPU with 24GB memory for sd3.5 fp16 bs1 lora training.

Installation

0. Microsoft Visual C++ Redistributable latest supported downloads

Please install Microsoft Visual C++ Redistributable if you have the following error:
0SError:[WinError 126] ... Error loading "...\Lib\site-packages\torch\lib\fbgemm dl1" or one of its dependencies.
https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170

1. Setup environment using setup.bat:

setup.bat include 5 steps:

• Step 1. Setup venv (optional)
• Step 2. Upgrade pip
• Step 3. Install torch (auto,optional) if setup venv install torch automatically.
• Step 4. Install other dependencies from requirements.txt
• Step 5. Download the Kolors Model Files (optional, must if you train Kolors)
• Step 6. Download the SD3.5 Large Model Files (optional, must if you train SD3.5 Large)
• Step 6. Download the SD3.5 Medium Model Files (optional, must if you train SD3.5 Medium)
  • Note: If you already have SD3.5 Large downloaded, you can only download the SD3.5 Medium transformer subfolder and replace it to the sd3.5 Large transformer subfolder.

git clone https://github.com/lrzjason/T2ITrainer.git
cd T2ITrainer
setup.bat

Goto point 2 after setup.bat

1. Setup environment manually (optional):

Clone repository

git clone https://github.com/lrzjason/T2ITrainer.git
cd T2ITrainer

Create virtual environment if you want:

python -m venv venv
call venv\Scripts\activate

If created virtual environment, install pytorch from https://pytorch.org/get-started/locally/ (Recommanded Torch version >= 2.3.0+cu121) Example:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Install dependencies:

pip install -r requirements.txt

Download Models seperately via huggingface-cli:

For Kolors:

huggingface-cli download Kwai-Kolors/Kolors --local-dir kolors_models/

For SD3.5 Large:

huggingface-cli download "stabilityai/stable-diffusion-3.5-large" --local-dir "sd3.5L/"

For SD3.5 Medium:

huggingface-cli download "stabilityai/stable-diffusion-3.5-medium" --local-dir "sd3.5M/"

2. Run the script (Kolors):

• PS: Kolors' original vae in fp16 would cause training error and produce black image when inference.
• Please download the fp16 fix vae from https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
• sdxl.vae.safetensors or sdxl_vae.safetensors and use one of them in vae_path option

python ui.py

2. Run the script (SD3.5 Large):

python ui_sd35.py

3. Testing:

• For kolors: Install plugins from:

  • ComfyUI-KwaiKolorsWrapper
  • ComfyUI-Kolors-MZ

  Then use the normal lora loader in ComfyUI.

4. Parameter Explanation:

Refer to my civitai article https://civitai.com/articles/7743

Star History

Change Logs:

• 2024-06-17: Initial UI release.
• 2024-06-25: Updated hunyuandit lora training script and UI for multiple training scripts.
• 2024-06-27: Fixed incorrect bucketing in hunyuandit lora training and added cache recreation option in UI.
• 2024-07-06: Fixed RGB to BGR channel issue in sd3 latent.
• 2024-07-08: Added merge_state_dict.py for merging Kolors and SDXL state dicts, saving as Kolors unet.
• 2024-07-11: Added Kolors training script to UI.
• 2024-07-11:(2) Added resume from checkpoint
• 2024-07-13: Remove some unused dependencies
• 2024-07-13:(2) Fix bugs on saved lora, added prefix to keys
• 2024-07-14: Avoid loading text encoder unnecessarily after cached latent
• 2024-07-15: Added setup.bat for windows users
• 2024-07-16: Added save and load config.json
• 2024-07-19: Fix model_path couldn't load the weight
• 2024-07-20: Adjust cache logic and fix caching bug. Please recreate cache after this update.
• 2024-07-20:(2) Add 2048 resolution, it could choose 1024, 2048 training based resolution now. It only recommanded to train 2048 resolution when you have >= 24GB vram.
• 2024-07-21: Update debiased estimation loss and change cache resize from default to area
• 2024-07-22: Fix when batch_size > 1, debiased estimation loss would cause error.
• 2024-07-22:(2) Thanks emojiiii for linux setup.sh script and related updates.
• 2024-07-28: Update caching logic, automatically update cache when added or deleted files.
• 2024-08-04: Add kolors slider training code and image generation code
• 2024-09-11: Merge dev branch. Fix debiased estimation loss and min-snr logic.
• 2024-09-14: Update captioner script: internvl2 and wd14
• 2024-09-17: Add more res buckets in 1024 scope
• 2024-09-30: T2Itrainer parameter explanation https://civitai.com/articles/7743
• 2024-09-30:(2) Fix small bug and remove 2048 training option.
• 2024-10-08: Add 512 resolution training
• 2024-10-23: Add sd3.5 Large training script
• 2024-10-24: Add freeze layers for SD3.5 training, default as 30 which would freeze 30~37 layers to keep texture
• 2024-10-30: Add model_path for SD3.5 training
• 2024-11-07: Modify sd3.5 freeze layers training
• 2025-01-02: Add flux lora training script. Thanks to diffusers and kohya codebase. Most code are referencied from them. (UI not ready yet. It is still in development.)

Acknowledgements:

• Thanks to minienglish1 and Freon in EveryDream Discord for the assistance.
• Special thanks to kohya ss for references from the training codebase.
• Thanks to Kblueleaf for coding reference on hunyuandit gradient checkpoint implementation.
• Thanks to Kolors for the open-source checkpoint.
• Thanks to comfyui for the wonderful codebase.
• Thanks to emojiiii for the setup.bat script and other updates.
• Thanks to Rohit Gandikota and related authors of Concept Sliders https://github.com/rohitgandikota/sliders