https://www.bilibili.com/video/BV1RbGTzfEQF/?vd_source=2ebab812a46c68cabe85acf4e3468677

可以自动断点继续炼

---

T2ITrainer is a diffusers based training script. It aims to provide simple yet implementation for lora training.

• 2025-07-16:
  add --config_path argument to train_flux_lora_ui_kontext.py support config file to start the kontext training

    python train_flux_lora_ui_kontext.py --config_path config.json

---

• PyTorch: torch>=2.3.0+cu121 (CUDA 12.1 supported)

---

Model Type	VRAM Requirements	Status
Kolors	11GB GPU	✅ Supported
SD3.5 (FP16 BS1)	24GB GPU	✅ Supported
Flux Fill,Kontext	24GB GPU	✅ Supported

---

❗ Mandatory: Install Microsoft Visual C++ Redistributable if encountering DLL errors

Recommended Method

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

• Handles: Virtual Environment • Dependency Installation • Model Downloads

Clone Repository 🌐

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

Virtual Environment 🛠️

    python -m venv venv
    call venv\Scripts\activate
    pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Model Downloads 📥

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

    # NF4 Flux Fill for low gpu
    huggingface-cli download "lrzjason/flux-fill-nf4" --local-dir flux_models/fill/

    # skip if downloaded nf 4 Flux Fill
    huggingface-cli download "black-forest-labs/FLUX.1-fill-dev" --local-dir flux_models/fill/
    
    # SD3.5 Models
    huggingface-cli download "stabilityai/stable-diffusion-3.5-large" --local-dir "sd3.5L/"

    # NF4 Flux kontext
    huggingface-cli download "lrzjason/flux-kontext-nf4" --local-dir flux_models/kontext/

Script	Command	Special Notes
Flux kontext	python ui_flux_fill.py	Requires diffusers>=0.32.0, 24GB VRAM Recommended
Flux Fill	python ui_flux_fill.py	Requires diffusers>=0.32.0, 24GB VRAM Recommended
Kolors	python ui.py	Needs Fixed VAE
SD3.5 Large	python ui_sd35.py	24GB VRAM Recommended

---

Inpainting Model Setup

  huggingface-cli download "lrzjason/flux-kontext-nf4" --local-dir flux_models/kontext/

For more details (example dataset):

• https://github.com/lrzjason/T2ITrainer/blob/main/doc/flux_kontext.md
• https://huggingface.co/datasets/lrzjason/object_removal_alpha_kontext

Inpainting Model Setup

  huggingface-cli download "lrzjason/flux-fill-nf4" --local-dir flux_models/fill/ 

For more details (example dataset):

• https://github.com/lrzjason/T2ITrainer/blob/main/doc/flux_fill.md
• https://huggingface.co/datasets/lrzjason/ObjectRemovalAlpha

Dev Model Installation

  huggingface-cli download "black-forest-labs/FLUX.1-dev" --local-dir flux_models/dev/

Category	Settings
Base Configuration	Rank 16, AdamW, Lr 1e-4
24GB GPU	512 resolution, Batch Size 1
VRAM Optimization	Use nf4 based training
Precision	bf16

Register WandB before using it Setup WandB env

pip install wandb
wandb login

Install Tensorboard first if you choice to use Tensorboard To visualize training data, run the following command in your terminal:

tensorboard --logdir=.\logs

• Kolors Workflow:

      # ComfyUI Plugins
      git clone https://github.com/kijai/ComfyUI-KwaiKolorsWrapper
      git clone https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ
  

• Configuration Guide: 📖 CivitAI Article

• Kolors Black Image Issue: Ensure you're using FP16 Fixed VAE
• VRAM Limitations: Adjust blocks_to_swap parameter (higher values reduce memory usage)
• Windows DLL Errors: Verify VC++ Redistributable installation

---

• https://github.com/lrzjason/T2ITrainer/blob/main/doc/change_logs.md

• 2025-07-20: Support to user original kontext repo to train lora. --exclude "*.safetensors" skip the big one model file which is not needed.

huggingface-cli download "black-forest-labs/FLUX.1-Kontext-dev" --local-dir flux_models/kontext_full/ --exclude "*.safetensors"

• 2025-07-16:
  add --config_path argument to train_flux_lora_ui_kontext.py support config file to start the kontext training

    python train_flux_lora_ui_kontext.py --config_path config.json

• Thanks to @sourceful support me making flux fill lora training script.

• 𝕏 Twitter: @Lrzjason
• 📧 Email: lrzjason@gmail.com
• 💬 QQ Group: 866612947
• 💬 WeChat ID: fkdeai
• 🎨 CivitAI: xiaozhijason

• Thanks to chenpipi0807 for Chinese translation and language switch support
• Thanks for diffusers and Terminus Research Group
• 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