Skip to content

Latest commit

 

History

History
362 lines (254 loc) · 13.4 KB

SD3.md

File metadata and controls

362 lines (254 loc) · 13.4 KB
Raw

Stable Diffusion 3

In this example, we'll be training a Stable Diffusion 3 model using the SimpleTuner toolkit and will be using the lora model type.

Prerequisites

Make sure that you have python installed; SimpleTuner does well with 3.10 or 3.11. Python 3.12 should not be used.

You can check this by running:

python --version

If you don't have python 3.11 installed on Ubuntu, you can try the following:

apt -y install python3.11 python3.11-venv

Container image dependencies

For Vast, RunPod, and TensorDock (among others), the following will work on a CUDA 12.2-12.4 image:

apt -y install nvidia-cuda-toolkit libgl1-mesa-glx

If libgl1-mesa-glx is not found, you might need to use libgl1-mesa-dri instead. Your mileage may vary.

Installation

Clone the SimpleTuner repository and set up the python venv:

git clone --branch=release https://github.com/bghira/SimpleTuner.git

cd SimpleTuner

python -m venv .venv

source .venv/bin/activate

pip install -U poetry pip

# Necessary on some systems to prevent it from deciding it knows better than us.
poetry config virtualenvs.create false

Depending on your system, you will run one of 3 commands:

# MacOS
poetry install -C install/apple

# Linux
poetry install

# Linux with ROCM
poetry install -C install/rocm

AMD ROCm follow-up steps

The following must be executed for an AMD MI300X to be useable:

apt install amd-smi-lib
pushd /opt/rocm/share/amd_smi
python3 -m pip install --upgrade pip
python3 -m pip install .
popd

Removing DeepSpeed & Bits n Bytes

These two dependencies cause numerous issues for container hosts such as RunPod and Vast.

To remove them after poetry has installed them, run the following command in the same terminal:

pip uninstall -y deepspeed bitsandbytes

Setting up the environment

To run SimpleTuner, you will need to set up a configuration file, the dataset and model directories, and a dataloader configuration file.

Configuration file

An experimental script, configure.py, may allow you to entirely skip this section through an interactive step-by-step configuration. It contains some safety features that help avoid common pitfalls.

Note: This doesn't configure your dataloader. You will still have to do that manually, later.

To run it:

python configure.py

⚠️ For users located in countries where Hugging Face Hub is not readily accessible, you should add HF_ENDPOINT=https://hf-mirror.com to your ~/.bashrc or ~/.zshrc depending on which $SHELL your system uses.

If you prefer to manually configure:

Copy config/config.json.example to config/config.json:

cp config/config.json.example config/config.json

There, you will need to modify the following variables:

{
  "model_type": "lora",
  "model_family": "sd3",
  "pretrained_model_name_or_path": "stabilityai/stable-diffusion-3.5-large",
  "output_dir": "/home/user/outputs/models",
  "validation_resolution": "1024x1024,1280x768",
  "validation_guidance": 3.0,
  "validation_prompt": "your main test prompt here",
  "user_prompt_library": "config/user_prompt_library.json"
}
  • pretrained_model_name_or_path - Set this to stabilityai/stable-diffusion-3.5-large. Note that you will need to log in to Huggingface and be granted access to download this model. We will go over logging in to Huggingface later in this tutorial.
    • If you prefer to train the older SD3.0 Medium (2B), use stabilityai/stable-diffusion-3-medium-diffusers instead.
  • MODEL_TYPE - Set this to lora.
  • MODEL_FAMILY - Set this to sd3.
  • OUTPUT_DIR - Set this to the directory where you want to store your checkpoints and validation images. It's recommended to use a full path here.
  • VALIDATION_RESOLUTION - As SD3 is a 1024px model, you can set this to 1024x1024.
    • Additionally, SD3 was fine-tuned on multi-aspect buckets, and other resolutions may be specified using commas to separate them: 1024x1024,1280x768
  • VALIDATION_GUIDANCE - SD3 benefits from a very-low value. Set this to 3.0.

There are a few more if using a Mac M-series machine:

  • mixed_precision should be set to no.

Quantised model training

Tested on Apple and NVIDIA systems, Hugging Face Optimum-Quanto can be used to reduce the precision and VRAM requirements well below the requirements of base SDXL training.

Inside your SimpleTuner venv:

pip install optimum-quanto

⚠️ If using a JSON config file, be sure to use this format in config.json instead of config.env:

{
  "base_model_precision": "int8-quanto",
  "text_encoder_1_precision": "no_change",
  "text_encoder_2_precision": "no_change",
  "text_encoder_3_precision": "no_change",
  "optimizer": "adamw_bf16"
}

For config.env users (deprecated):

# choices: int8-quanto, int4-quanto, int2-quanto, fp8-quanto
# int8-quanto was tested with a single subject dreambooth LoRA.
# fp8-quanto does not work on Apple systems. you must use int levels.
# int2-quanto is pretty extreme and gets the whole rank-1 LoRA down to about 13.9GB VRAM.
# may the gods have mercy on your soul, should you push things Too Far.
export TRAINER_EXTRA_ARGS="--base_model_precision=int8-quanto"

# Maybe you want the text encoders to remain full precision so your text embeds are cake.
# We unload the text encoders before training, so, that's not an issue during training time - only during pre-caching.
# Alternatively, you can go ham on quantisation here and run them in int4 or int8 mode, because no one can stop you.
export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --text_encoder_1_precision=no_change --text_encoder_2_precision=no_change"

# When you're quantising the model, --base_model_default_dtype is set to bf16 by default. This setup requires adamw_bf16, but saves the most memory.
# adamw_bf16 only supports bf16 training, but any other optimiser will support both bf16 or fp32 training precision.
export OPTIMIZER="adamw_bf16"

Dataset considerations

It's crucial to have a substantial dataset to train your model on. There are limitations on the dataset size, and you will need to ensure that your dataset is large enough to train your model effectively. Note that the bare minimum dataset size is TRAIN_BATCH_SIZE * GRADIENT_ACCUMULATION_STEPS as well as more than VAE_BATCH_SIZE. The dataset will not be useable if it is too small.

Depending on the dataset you have, you will need to set up your dataset directory and dataloader configuration file differently. In this example, we will be using pseudo-camera-10k as the dataset.

In your /home/user/simpletuner/config directory, create a multidatabackend.json:

[
  {
    "id": "pseudo-camera-10k-sd3",
    "type": "local",
    "crop": true,
    "crop_aspect": "square",
    "crop_style": "center",
    "resolution": 1.0,
    "minimum_image_size": 0,
    "maximum_image_size": 1.0,
    "target_downsample_size": 1.0,
    "resolution_type": "area",
    "cache_dir_vae": "cache/vae/sd3/pseudo-camera-10k",
    "instance_data_dir": "/home/user/simpletuner/datasets/pseudo-camera-10k",
    "disabled": false,
    "skip_file_discovery": "",
    "caption_strategy": "filename",
    "metadata_backend": "discovery"
  },
  {
    "id": "text-embeds",
    "type": "local",
    "dataset_type": "text_embeds",
    "default": true,
    "cache_dir": "cache/text/sd3/pseudo-camera-10k",
    "disabled": false,
    "write_batch_size": 128
  }
]

Then, create a datasets directory:

mkdir -p datasets
pushd datasets
    huggingface-cli download --repo-type=dataset bghira/pseudo-camera-10k --local-dir=pseudo-camera-10k
popd

This will download about 10k photograph samples to your datasets/pseudo-camera-10k directory, which will be automatically created for you.

Login to WandB and Huggingface Hub

You'll want to login to WandB and HF Hub before beginning training, especially if you're using push_to_hub: true and --report_to=wandb.

If you're going to be pushing items to a Git LFS repository manually, you should also run git config --global credential.helper store

Run the following commands:

wandb login

and

huggingface-cli login

Follow the instructions to log in to both services.

Executing the training run

From the SimpleTuner directory, one simply has to run:

bash train.sh

This will begin the text embed and VAE output caching to disk.

For more information, see the dataloader and tutorial documents.

Notes & troubleshooting tips

Skip-layer guidance (SD3.5 Medium)

StabilityAI recommends enabling SLG (Skip-layer guidance) on SD 3.5 Medium inference. This doesn't impact training results, only the validation sample quality.

The following values are recommended for config.json:

{
  "--validation_guidance_skip_layers": [7, 8, 9],
  "--validation_guidance_skip_layers_start": 0.01,
  "--validation_guidance_skip_layers_stop": 0.2,
  "--validation_guidance_skip_scale": 2.8,
  "--validation_guidance": 4.0,
  "--flux_use_uniform_schedule": true,
  "--flux_schedule_auto_shift": true
}
  • ..skip_scale determines how much to scale the positive prompt prediction during skip-layer guidance. The default value of 2.8 is safe for the base model's skip value of 7, 8, 9 but will need to be increased if more layers are skipped, doubling it for each additional layer.
  • ..skip_layers tells which layers to skip during the negative prompt prediction.
  • ..skip_layers_start determine the fraction of the inference pipeline during which skip-layer guidance should begin to be applied.
  • ..skip_layers_stop will set the fraction of the total number of inference steps after which SLG will no longer be applied.

SLG can be applied for fewer steps for a weaker effect or less reduction of inference speed.

It seems that extensive training of a LoRA or LyCORIS model will require modification to these values, though it's not clear how exactly it changes.

Lower CFG must be used during inference.

Model instability

The SD 3.5 Large 8B model has potential instabilities during training:

  • High --max_grad_norm values will allow the model to explore potentially dangerous weight updates
  • Learning rates can be extremely sensitive; 1e-5 works with StableAdamW but 4e-5 could explode
  • Higher batch sizes help a lot
  • The stability is not impacted by disabling quantisation or training in pure fp32

Official training code was not released alongside SD3.5, leaving developers to guess how to implement the training loop based on the SD3.5 repository contents.

Some changes were made to SimpleTuner's SD3.5 support:

  • Excluding more layers from quantisation
  • No longer zeroing T5 padding space by default (--t5_padding)
  • Offering a switch (--sd3_clip_uncond_behaviour and --sd3_t5_uncond_behaviour) to use empty encoded blank captions for unconditional predictions (empty_string, default) or zeros (zero), not a recommended setting to tweak.
  • SD3.5 training loss function was updated to match that found in the upstream StabilityAI/SD3.5 repository
  • Updated default --flux_schedule_shift value to 3 to match the static 1024px value for SD3
    • StabilityAI followed-up with documentation to use --flux_schedule_shift=1 with --flux_use_uniform_schedule
    • Community members have reported that --flux_schedule_auto_shift works better when using mult-aspect or multi-resolution training
  • Updated the hard-coded tokeniser sequence length limit to 256 with the option to revert it to 77 tokens to save disk space or compute at the cost of output quality degradation

Stable configuration values

These options have been known to keep SD3.5 in-tact for as long as possible:

  • optimizer=adamw_bf16
  • flux_schedule_shift=1
  • learning_rate=1e-4
  • batch_size=4 * 3 GPUs
  • max_grad_norm=0.1
  • base_model_precision=int8-quanto
  • No loss masking or dataset regularisation, as their contribution to this instability is unknown
  • validation_guidance_skip_layers=[7,8,9]

Lowest VRAM config

  • OS: Ubuntu Linux 24
  • GPU: A single NVIDIA CUDA device (10G, 12G)
  • System memory: 50G of system memory approximately
  • Base model precision: bnb-nf4
  • Optimiser: Lion 8Bit Paged, bnb-lion8bit-paged
  • Resolution: 512px
  • Batch size: 1, zero gradient accumulation steps
  • DeepSpeed: disabled / unconfigured
  • PyTorch: 2.5

SageAttention

When using --attention_mechanism=sageattention, inference can be sped-up at validation time.

Note: This isn't compatible with every model configuration, but it's worth trying.

Masked loss

If you are training a subject or style and would like to mask one or the other, see the masked loss training section of the Dreambooth guide.

Regularisation data

For more information on regularisation datasets, see this section and this section of the Dreambooth guide.

Quantised training

See this section of the Dreambooth guide for information on configuring quantisation for SD3 and other models.

CLIP score tracking

If you wish to enable evaluations to score the model's performance, see this document for information on configuring and interpreting CLIP scores.