### Manual Installation and Launch of Flux Gym

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Steps to clone the repository, set up a virtual environment, install dependencies (including sd-scripts and PyTorch), and launch the Flux Gym application. Ensure to activate the correct virtual environment and install sd-scripts dependencies before Flux Gym's.

```bash
# Clone fluxgym and the Kohya sd-scripts backend
git clone https://github.com/cocktailpeanut/fluxgym
cd fluxgym
git clone -b sd3 https://github.com/kohya-ss/sd-scripts

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

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

# Install Flux Gym dependencies
pip install -r requirements.txt

# Install PyTorch Nightly for CUDA 12.1 (standard GPUs)
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

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

# Launch the app
python app.py
# Open http://localhost:7860 in your browser
```

--------------------------------

### Install Flux Gym Dependencies

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Return to the root Flux Gym directory and install its dependencies into the activated virtual environment.

```bash
cd ..
pip install -r requirements.txt
```

--------------------------------

### Install Kohya Scripts Dependencies

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Navigate to the sd-scripts directory and install its dependencies into the activated virtual environment.

```bash
cd sd-scripts
pip install -r requirements.txt
```

--------------------------------

### Start Flux Gym Application

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Run the Flux Gym application from the root directory with the virtual environment activated. This launches the web UI.

```bash
python app.py
```

--------------------------------

### Docker Installation and Launch of Flux Gym

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Instructions for setting up Flux Gym using Docker Compose. This method simplifies dependency management and deployment. Optionally, user and group IDs can be overridden.

```bash
git clone https://github.com/cocktailpeanut/fluxgym
cd fluxgym
git clone -b sd3 https://github.com/kohya-ss/sd-scripts

# Optional: override user/group IDs (default 1000)
# export PUID=$(id -u) PGID=$(id -g)

docker compose up -d --build
# Open http://localhost:7860
```

--------------------------------

### Build and Run Flux Gym with Docker Compose

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Build the Docker image and start the Flux Gym application in detached mode using Docker Compose. Access the UI via http://localhost:7860.

```bash
docker compose up -d --build
```

--------------------------------

### Install PyTorch Nightly (cu121)

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Install the nightly build of PyTorch with CUDA 12.1 support. This is a common requirement for GPU acceleration.

```bash
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
```

--------------------------------

### Configuring Base Models in models.yaml

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Example YAML configuration for defining base models available in Flux Gym. Each entry specifies the Hugging Face repository, base model, license, and file name. Custom models can be added by following this structure.

```yaml
# models.yaml — add custom base models here

# Built-in: Flux1-dev (non-commercial license)
flux-dev:
    repo: cocktailpeanut/xulf-dev
    base: black-forest-labs/FLUX.1-dev
    license: other
    license_name: flux-1-dev-non-commercial-license
    license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
    file: flux1-dev.sft

# Built-in: Flux1-schnell (Apache 2.0)
flux-schnell:
    repo: black-forest-labs/FLUX.1-schnell
    base: black-forest-labs/FLUX.1-schnell
    license: apache-2.0
    file: flux1-schnell.safetensors

# Built-in: Flux1-dev2pro (improved training stability)
bdsqlsz/flux1-dev2pro-single:
    repo: bdsqlsz/flux1-dev2pro-single
    base: black-forest-labs/FLUX.1-dev
    license: other
    license_name: flux-1-dev-non-commercial-license
    license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
    file: flux1-dev2pro.safetensors

# Example: adding a custom community model
my-custom-flux:
    repo: myuser/my-flux-variant
    base: black-forest-labs/FLUX.1-dev
    license: apache-2.0
    file: my-flux-variant.safetensors
```

--------------------------------

### Clone Flux Gym and Kohya Scripts Repositories

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Clone the Flux Gym repository and the specific sd3 branch of Kohya Scripts. This is the initial step for manual installation.

```bash
git clone https://github.com/cocktailpeanut/fluxgym
cd fluxgym
git clone -b sd3 https://github.com/kohya-ss/sd-scripts
```

--------------------------------

### Install PyTorch Nightly (cu128) and Update Bitsandbytes

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Install the nightly build of PyTorch with CUDA 12.8 support and update bitsandbytes for compatibility with NVIDIA RTX 50-series GPUs.

```bash
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
pip install -U bitsandbytes
```

--------------------------------

### Automatic Model Download Function

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Python code demonstrating the internal `download(base_model)` function used by Flux Gym. This function automatically downloads necessary model components from Hugging Face to the local `models/` directory, skipping existing files. It's triggered before training starts.

```python
# Internal usage within app.py — triggered automatically on training start
# Model files are placed in:
#   models/unet/         → flux1-dev.sft (or variant)
#   models/vae/          → ae.sft
#   models/clip/         → clip_l.safetensors, t5xxl_fp16.safetensors

download("flux-dev")
# Console output:
# download flux-dev
# downloading ae.sft...
# download clip_l.safetensors
# download t5xxl_fp16.safetensors

# Resulting directory layout:
# models/
#   unet/flux1-dev.sft
#   vae/ae.sft
#   clip/clip_l.safetensors
#   clip/t5xxl_fp16.safetensors
```

--------------------------------

### start_training

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Orchestrates the full training pipeline, including directory creation, model download, script writing, training execution, and auto-generation of a Hugging Face model card.

```APIDOC
## `start_training(base_model, lora_name, train_script, train_config, sample_prompts)` — Training Execution

Orchestrates the full training pipeline: creates output directories, triggers model download, writes the shell script and dataset TOML to disk, executes training via `bash` (or the `.bat` equivalent on Windows) with streamed log output, then auto-generates a Hugging Face model card README upon completion.

```python
# Called by the Gradio UI when the user clicks "Start training"
# This is a generator — yields log lines to the terminal LogsView component

for log_line in start_training(
    base_model="flux-dev",
    lora_name="My Persian Miniature Style",
    train_script=sh_script,      # output of gen_sh()
    train_config=toml_config,    # output of gen_toml()
    sample_prompts="prsmn painting of a cat --d 42\nprsmn painting of a landscape --d 42",
):
    print(log_line)

# Side effects on disk after completion:
# outputs/my-persian-miniature-style/
#   train.sh               ← generated training script
#   dataset.toml           ← Kohya dataset config
#   sample_prompts.txt     ← sample image prompts
#   my-persian-miniature-style-000004.safetensors   ← checkpoint at epoch 4
#   my-persian-miniature-style-000008.safetensors   ← checkpoint at epoch 8
#   my-persian-miniature-style.safetensors          ← final LoRA weights
#   sample/                ← generated sample images (if configured)
#   README.md              ← auto-generated HuggingFace model card
```
```

--------------------------------

### Prepare Training Dataset

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Copies uploaded images into the training dataset folder, resizes them to the configured resolution (maintaining aspect ratio via LANCZOS resampling), and writes paired .txt caption files. Skips writing caption files that already exist.

```python
# Called just before start_training() — assembles the on-disk dataset
destination_folder = "datasets/my-lora"
size = 512  # target resolution (images are resized to this on their shorter side)

# inputs[0] = list of image file paths
# inputs[1..N] = corresponding caption strings
images = ["/tmp/img0.png", "/tmp/img1.png"]
captions = ["p3rs0n riding a bike", "p3rs0n as a bodybuilder"]

result = create_dataset(destination_folder, size, images, *captions)
# result = "datasets/my-lora"
# Creates:
#   datasets/my-lora/img0.png   (resized to 512px on short side)
#   datasets/my-lora/img0.txt   → "p3rs0n riding a bike"
#   datasets/my-lora/img1.png
#   datasets/my-lora/img1.txt   → "p3rs0n as a bodybuilder"
```

--------------------------------

### create_dataset

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Prepares the training dataset by copying and resizing images, and creating corresponding `.txt` caption files. It preserves existing caption files, ensuring manual annotations are not overwritten.

```APIDOC
## `create_dataset(destination_folder, size, *inputs)` — Dataset Preparation

Copies uploaded images into the training dataset folder, resizes them to the configured resolution (maintaining aspect ratio via LANCZOS resampling), and writes paired `.txt` caption files. Skips writing caption files that already exist (preserving manually uploaded `.txt` captions).

```python
# Called just before start_training() — assembles the on-disk dataset
destination_folder = "datasets/my-lora"
size = 512  # target resolution (images are resized to this on their shorter side)

# inputs[0] = list of image file paths
# inputs[1..N] = corresponding caption strings
images = ["/tmp/img0.png", "/tmp/img1.png"]
captions = ["p3rs0n riding a bike", "p3rs0n as a bodybuilder"]

result = create_dataset(destination_folder, size, images, *captions)
# result = "datasets/my-lora"
# Creates:
#   datasets/my-lora/img0.png   (resized to 512px on short side)
#   datasets/my-lora/img0.txt   → "p3rs0n riding a bike"
#   datasets/my-lora/img1.png
#   datasets/my-lora/img1.txt   → "p3rs0n as a bodybuilder"
```
```

--------------------------------

### Sample Image Generation Flags

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Utilize various flags to control sample image generation during training. These include negative prompts (--n), dimensions (--w, --h), seed (--d), CFG scale (--l), and steps (--s). Prompt weighting syntax is also supported.

```text
--n: Negative prompt up to the next option.
--w: Specifies the width of the generated image.
--h: Specifies the height of the generated image.
--d: Specifies the seed of the generated image.
--l: Specifies the CFG scale of the generated image.
--s: Specifies the number of steps in the generation.
```

--------------------------------

### gen_sh

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Assembles the complete `accelerate launch` shell command for Kohya's training script. It automatically adapts the optimizer and memory strategy based on the specified VRAM tier and appends advanced Kohya flags.

```APIDOC
## `gen_sh(...)` — Training Shell Script Generation

Assembles the complete `accelerate launch` shell command for Kohya's `flux_train_network.py`, adapting the optimizer and memory strategy to the selected VRAM tier. For 12GB uses split-mode + single-block training; for 16GB uses Adafactor without split; for 20GB+ uses AdamW8bit. Advanced Kohya flags changed from their defaults are appended automatically.

```python
sh_script = gen_sh(
    base_model="flux-dev",
    output_name="my-lora",
    resolution=512,
    seed=42,
    workers=2,
    learning_rate="8e-4",
    network_dim=4,          # LoRA rank
    max_train_epochs=16,
    save_every_n_epochs=4,
    timestep_sampling="shift",
    guidance_scale=1.0,
    vram="16G",             # "12G", "16G", or "20G"
    sample_prompts="p3rs0n riding a bike --d 42",
    sample_every_n_steps=100,
    # *advanced_components omitted (uses defaults)
)
# Produces a script like:
# accelerate launch \
#   --mixed_precision bf16 \
#   --num_cpu_threads_per_process 1 \
#   sd-scripts/flux_train_network.py \
#   --pretrained_model_name_or_path "/abs/path/models/unet/flux1-dev.sft" \
#   --clip_l "/abs/path/models/clip/clip_l.safetensors" \
#   --t5xxl "/abs/path/models/clip/t5xxl_fp16.safetensors" \
#   --ae "/abs/path/models/vae/ae.sft" \
#   --optimizer_type adafactor \
#   --optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False" \
#   --lr_scheduler constant_with_warmup \
#   --max_grad_norm 0.0 \
#   --sample_prompts="/abs/path/outputs/my-lora/sample_prompts.txt" \
#   --sample_every_n_steps="100" \
#   --network_module networks.lora_flux \
#   --network_dim 4 \
#   --learning_rate 8e-4 \
#   --max_train_epochs 16 \
#   --save_every_n_epochs 4 \
#   --timestep_sampling shift \
#   --guidance_scale 1.0 \
#   --loss_type l2
```
```

--------------------------------

### Generate Training Shell Script

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Assembles the complete `accelerate launch` shell command for Kohya's `flux_train_network.py`. Adapts the optimizer and memory strategy to the selected VRAM tier (12GB, 16GB, or 20GB+). Advanced Kohya flags are appended automatically.

```python
sh_script = gen_sh(
    base_model="flux-dev",
    output_name="my-lora",
    resolution=512,
    seed=42,
    workers=2,
    learning_rate="8e-4",
    network_dim=4,          # LoRA rank
    max_train_epochs=16,
    save_every_n_epochs=4,
    timestep_sampling="shift",
    guidance_scale=1.0,
    vram="16G",             # "12G", "16G", or "20G"
    sample_prompts="p3rs0n riding a bike --d 42",
    sample_every_n_steps=100,
    # *advanced_components omitted (uses defaults)
)
# Produces a script like:
# accelerate launch \
#   --mixed_precision bf16 \
#   --num_cpu_threads_per_process 1 \
#   sd-scripts/flux_train_network.py \
#   --pretrained_model_name_or_path "/abs/path/models/unet/flux1-dev.sft" \
#   --clip_l "/abs/path/models/clip/clip_l.safetensors" \
#   --t5xxl "/abs/path/models/clip/t5xxl_fp16.safetensors" \
#   --ae "/abs/path/models/vae/ae.sft" \
#   --optimizer_type adafactor \
#   --optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False" \
#   --lr_scheduler constant_with_warmup \
#   --max_grad_norm 0.0 \
#   --sample_prompts="/abs/path/outputs/my-lora/sample_prompts.txt" \
#   --sample_every_n_steps="100" \
#   --network_module networks.lora_flux \
#   --network_dim 4 \
#   --learning_rate 8e-4 \
#   --max_train_epochs 16 \
#   --save_every_n_epochs 4 \
#   --timestep_sampling shift \
#   --guidance_scale 1.0 \
#   --loss_type l2
```

--------------------------------

### Activate Python Virtual Environment (Linux)

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Create and activate a Python virtual environment on Linux. This isolates project dependencies.

```bash
python -m venv env
source env/bin/activate
```

--------------------------------

### Activate Python Virtual Environment (Windows)

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Create and activate a Python virtual environment on Windows. This isolates project dependencies.

```bash
python -m venv env
env\Scripts\activate
```

--------------------------------

### Uploading Pre-Made Caption Files

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Include `.txt` caption files with matching base names alongside image files for automatic detection. This skips AI captioning for those specific images. Ensure the caption file's base name exactly matches its corresponding image file.

```text
# File naming convention — caption file must share the exact base name as its image:
img0.png  →  img0.txt
img1.jpg  →  img1.txt
photo_42.webp  →  photo_42.txt

# Example caption file contents (img0.txt):
p3rs0n standing in front of the Eiffel Tower, wearing a blue coat, sunny day

# Upload both image and txt files together via the Gradio file picker.
# The caption text will pre-populate in the caption field for that image.
# If no .txt file is found for an image, the trigger word is used as the default.
```

--------------------------------

### Advanced Sample Image Prompts with Seed

Source: https://github.com/cocktailpeanut/fluxgym/blob/main/README.md

Use advanced flags like --d to specify the seed for consistent sample image generation, allowing visualization of LoRA evolution. This is useful for debugging and understanding training progress.

```text
hrld person is riding a bike --d 42
hrld person is a body builder --d 42
hrld person is a rock star --d 42
```

--------------------------------

### Sample Image Prompt Flags for Kohya's sd-scripts

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Use inline flags within prompts to control sample image generation during training. Flags like `--d` fix the seed for consistent evolution tracking. Refer to Kohya's sd-scripts documentation for a full flag reference.

```text
# Basic sample prompts (one per line in the "Sample Image Prompts" field)
p3rs0n is riding a bike
p3rs0n is a body builder
p3rs0n is a rock star

# With advanced flags — use --d to fix the seed so you can watch the LoRA evolve
p3rs0n is riding a bike --d 42
p3rs0n is a body builder --d 42
p3rs0n is a rock star --d 42

# Full flag reference:
# --n  Negative prompt (e.g.: --n blurry, low quality)
# --w  Image width in pixels  (e.g.: --w 768)
# --h  Image height in pixels (e.g.: --h 1024)
# --d  Seed                   (e.g.: --d 42)
# --l  CFG scale              (e.g.: --l 3.5)
# --s  Number of steps        (e.g.: --s 28)

# Combined example
p3rs0n portrait in oil painting style --d 42 --w 768 --h 1024 --s 28 --n ugly, blurry
```

--------------------------------

### Execute LoRA Training Pipeline

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Orchestrates the full training pipeline, including directory creation, model download, script generation, and execution. Yields log lines for real-time feedback. Ensure necessary scripts and configurations are generated prior to calling.

```python
# Called by the Gradio UI when the user clicks "Start training"
# This is a generator — yields log lines to the terminal LogsView component

for log_line in start_training(
    base_model="flux-dev",
    lora_name="My Persian Miniature Style",
    train_script=sh_script,      # output of gen_sh()
    train_config=toml_config,    # output of gen_toml()
    sample_prompts="prsmn painting of a cat --d 42\nprsmn painting of a landscape --d 42",
):
    print(log_line)

# Side effects on disk after completion:
# outputs/my-persian-miniature-style/
#   train.sh               ← generated training script
#   dataset.toml           ← Kohya dataset config
#   sample_prompts.txt     ← sample image prompts
#   my-persian-miniature-style-000004.safetensors   ← checkpoint at epoch 4
#   my-persian-miniature-style-000008.safetensors   ← checkpoint at epoch 8
#   my-persian-miniature-style.safetensors          ← final LoRA weights
#   sample/                ← generated sample images (if configured)
#   README.md              ← auto-generated HuggingFace model card
```

--------------------------------

### run_captioning

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Generates detailed captions for training images using the multimodalart/Florence-2-large-no-flash-attn model. The concept sentence is prepended to each caption, and the model is loaded on-demand to manage VRAM.

```APIDOC
## `run_captioning(images, concept_sentence, *captions)` — AI Auto-Captioning

Uses the `multimodalart/Florence-2-large-no-flash-attn` vision-language model to generate detailed captions for uploaded training images. The concept sentence (trigger word) is automatically prepended to each generated caption. The model is loaded on-demand to GPU and unloaded after completion to free VRAM for training.

```python
# Triggered when the user clicks "Add AI captions with Florence-2" in the UI
# images: list of file paths to uploaded images
# concept_sentence: trigger word, e.g. "p3rs0n"
# captions: current caption values (can be empty strings)

images = ["/tmp/img0.png", "/tmp/img1.png"]
concept_sentence = "p3rs0n"
captions = ["", ""]

# Generator — yields updated caption list after each image is processed
for updated_captions in run_captioning(images, concept_sentence, *captions):
    print(updated_captions)
# Output after img0.png:
# ["p3rs0n a person wearing a red jacket standing in a park", ""]
# Output after img1.png:
# ["p3rs0n a person wearing a red jacket standing in a park",
#  "p3rs0n a close-up portrait of a person with curly hair"]
```
```

--------------------------------

### Hugging Face Login and LoRA Upload

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Authenticates with Hugging Face using a token, saving it locally for subsequent uploads. Uploads the trained LoRA folder, including weights and README, to a specified Hugging Face repository. Ensure the LoRA output path is correct and repository details are provided.

```python
# Step 1: authenticate (triggered by "Login" button)
# Token is saved to ./HF_TOKEN — never sent anywhere except Hugging Face
login_hf("hf_yourTokenHere...")
# On success: saves token, returns updated UI visibility states

# Step 2: publish trained LoRA (triggered by "Upload to HuggingFace" button)
upload_hf(
    base_model="flux-dev",
    lora_rows="/abs/path/outputs/my-persian-miniature-style",
    repo_owner="myusername",
    repo_name="my-persian-miniature-style-lora",
    repo_visibility="public",   # or "private"
    hf_token="hf_yourTokenHere..."
)
# Uploads all files in the LoRA output folder to:
# https://huggingface.co/myusername/my-persian-miniature-style-lora
# Displays: [Upload Complete] https://huggingface.co/myusername/my-persian-miniature-style-lora
```

--------------------------------

### Generate Dataset Configuration (TOML)

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Generates the TOML dataset configuration string consumed by Kohya's training script. Configures shuffling, caption extension, batch size, resolution, and the image subset with its trigger token and repeat count.

```python
toml_config = gen_toml(
    dataset_folder="datasets/my-lora",
    resolution=512,
    class_tokens="p3rs0n",
    num_repeats=10
)
print(toml_config)
# [general]
# shuffle_caption = false
# caption_extension = '.txt'
# keep_tokens = 1
#
# [[datasets]]
# resolution = 512
# batch_size = 1
# keep_tokens = 1
#
#   [[datasets.subsets]]
#   image_dir = '/absolute/path/to/datasets/my-lora'
#   class_tokens = 'p3rs0n'
#   num_repeats = 10
```

--------------------------------

### Generate AI Captions with Florence-2

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Uses the multimodalart/Florence-2-large-no-flash-attn model to generate detailed captions for training images. The concept sentence is prepended to each caption. The model is loaded on-demand to GPU and unloaded after completion.

```python
# Triggered when the user clicks "Add AI captions with Florence-2" in the UI
# images: list of file paths to uploaded images
# concept_sentence: trigger word, e.g. "p3rs0n"
# captions: current caption values (can be empty strings)

images = ["/tmp/img0.png", "/tmp/img1.png"]
concept_sentence = "p3rs0n"
captions = ["", ""]

# Generator — yields updated caption list after each image is processed
for updated_captions in run_captioning(images, concept_sentence, *captions):
    print(updated_captions)
# Output after img0.png:
# ["p3rs0n a person wearing a red jacket standing in a park", ""]
# Output after img1.png:
# ["p3rs0n a person wearing a red jacket standing in a park",
#  "p3rs0n a close-up portrait of a person with curly hair"]
```

--------------------------------

### gen_toml

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Generates the TOML configuration string required by Kohya's training script. It sets parameters such as resolution, caption extension, and defines the image subset with its trigger token and repeat count.

```APIDOC
## `gen_toml(dataset_folder, resolution, class_tokens, num_repeats)` — Dataset Config Generation

Generates the TOML dataset configuration string consumed by Kohya's training script. Configures shuffling, caption extension, batch size, resolution, and the image subset with its trigger token and repeat count.

```python
toml_config = gen_toml(
    dataset_folder="datasets/my-lora",
    resolution=512,
    class_tokens="p3rs0n",
    num_repeats=10
)
print(toml_config)
# [general]
# shuffle_caption = false
# caption_extension = '.txt'
# keep_tokens = 1
#
# [[datasets]]
# resolution = 512
# batch_size = 1
# keep_tokens = 1
#
#   [[datasets.subsets]]
#   image_dir = '/absolute/path/to/datasets/my-lora'
#   class_tokens = 'p3rs0n'
#   num_repeats = 10
```
```

--------------------------------

### login_hf and upload_hf

Source: https://context7.com/cocktailpeanut/fluxgym/llms.txt

Handles Hugging Face authentication and publishing of trained LoRA models. `login_hf` validates and saves the token, while `upload_hf` pushes the model to a specified Hugging Face repository.

```APIDOC
## `login_hf(hf_token)` / `upload_hf(...)` — Hugging Face Publishing

`login_hf` validates a Hugging Face token via the `HfApi`, saves it to a local `HF_TOKEN` file, and updates the UI to show the authenticated account. `upload_hf` uses `huggingface_util.upload` from Kohya to push the selected trained LoRA folder (including safetensors weights and auto-generated README) to a new or existing Hugging Face repository.

```python
# Step 1: authenticate (triggered by "Login" button)
# Token is saved to ./HF_TOKEN — never sent anywhere except Hugging Face
login_hf("hf_yourTokenHere...")
# On success: saves token, returns updated UI visibility states

# Step 2: publish trained LoRA (triggered by "Upload to HuggingFace" button)
upload_hf(
    base_model="flux-dev",
    lora_rows="/abs/path/outputs/my-persian-miniature-style",
    repo_owner="myusername",
    repo_name="my-persian-miniature-style-lora",
    repo_visibility="public",   # or "private"
    hf_token="hf_yourTokenHere..."
)
# Uploads all files in the LoRA output folder to:
# https://huggingface.co/myusername/my-persian-miniature-style-lora
# Displays: [Upload Complete] https://huggingface.co/myusername/my-persian-miniature-style-lora
```
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.