Flux Gym

To install manually, you must clone both the fluxgym repository and the sd-scripts repository (specifically the sd3 branch), set up a Python virtual environment, and install the required dependencies for both.

# 1. Clone repositories
git clone https://github.com/cocktailpeanut/fluxgym
cd fluxgym
git clone -b sd3 https://github.com/kohya-ss/sd-scripts

# 2. Create and activate virtual environment
# Windows:
python -m venv env
env\Scripts\activate
# Linux:
python -m venv env
source env/bin/activate

# 3. Install dependencies
cd sd-scripts
pip install -r requirements.txt
cd ..
pip install -r requirements.txt

# 4. Install PyTorch Nightly
# Standard:
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# For NVIDIA RTX 50-series (e.g., 5090):
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
pip install -U bitsandbytes

# 5. Start the application
python app.py

The Flux Gym WebUI provides several configuration options for training FLUX LoRAs. Key settings include:

• LoRA Info:

  • lora_name: A unique name for your LoRA.
  • concept_sentence: The trigger word or sentence (e.g., p3rs0n or in the style of CNSTLL).
  • base_model: The foundation model (can be extended by editing models.yaml).
  • vram: Target VRAM capacity (20G, 16G, or 12G).
  • num_repeats: Number of repeat trains per image.
  • max_train_epochs: Maximum training epochs.
  • sample_prompts: Text prompts for generating sample images (separated by new lines).
  • sample_every_n_steps: Frequency of sample image generation.
  • resolution: The target resize resolution for dataset images.

• Advanced options:

  • --seed: Random seed for reproducibility.
  • --max_data_loader_n_workers: Number of data loader workers.
  • --learning_rate: The training learning rate (default: 8e-4).
  • --save_every_n_epochs: Frequency of saving checkpoints.
  • --guidance_scale: Guidance scale value.
  • --timestep_sampling: Timestep sampling method (e.g., shift).
  • --network_dim: LoRA Rank (range: 4 to 128, step 4).

You can deploy Fluxgym using Docker Compose. The configuration maps port 7860 to the host and requires an NVIDIA GPU.

Important CUDA Configuration: If you are running CUDA 12.4 drivers, you must change the dockerfile setting in the build section from Dockerfile to Dockerfile.cuda12.4. Otherwise, leave it as Dockerfile.

Environment Variables:

• PUID: User ID for the container (defaults to 1000).
• PGID: Group ID for the container (defaults to 1000).

Volumes:

• /etc/localtime and /etc/timezone are mounted as read-only to sync container time with the host.
• The current directory ./ is mounted to /app/fluxgym inside the container.

services:

fluxgym:
    build:
      context: .
      # change the dockerfile to Dockerfile.cuda12.4 if you are running CUDA 12.4 drivers otherwise leave as is
      dockerfile: Dockerfile
    image: fluxgym
    container_name: fluxgym
    ports:
      - 7860:7860
    environment:
      - PUID=${PUID:-1000}
      - PGID=${PGID:-1000}
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /etc/timezone:/etc/timezone:ro
      - ./:/app/fluxgym
    stop_signal: SIGKILL
    tty: true
    deploy:
      resources:
        reservations:
          devices:
          - driver: nvidia
            count: all
            capabilities: [gpu]
    restart: unless-stopped