README.md

July 22, 2026 ยท View on GitHub

Example configurations

These configurations are provided as an easy way to immediately run a training session with SimpleTuner across a large number of architectures.

The options are set up so that a 24G card (NVIDIA 4090) can run training out of the box. In order to do this, compromises were made for resolution, training batch size, or LoRA rank.

It's recommended to use these only as a basic starting point.

Running an example

All examples can be easily run without modifying the configurations.

We'll assume you don't have any python dependencies installed yet, and that an NVIDIA device is in use.

To run kontext.peft-lora example:

Option 1 (Recommended - pip install):

pip install 'simpletuner[cuda]'
# or for CUDA 13 / Blackwell users: pip install 'simpletuner[cuda13]' --extra-index-url https://download.pytorch.org/whl/cu130
simpletuner train example=kontext.peft-lora

Option 2 (Git clone method):

git clone https://github.com/bghira/simpletuner
cd simpletuner
python3.13 -m venv .venv
. .venv/bin/activate
pip install -e .
simpletuner train env=examples/kontext.peft-lora

Option 3 (Legacy method - still works):

git clone https://github.com/bghira/simpletuner
cd simpletuner
python3.13 -m venv .venv
. .venv/bin/activate
pip install -e .
ENV=examples/kontext.peft-lora ./train.sh

This will automatically download an example reference dataset, pre-cache embeds, and run 100 steps of training on a standard PEFT LoRA.

ACE-Step examples are split by model generation:

  • ace_step-v1-0.peft-lora for the original ACE-Step v1 3.5B path
  • ace_step-v1-5.peft-lora for the forward-compatible ACE-Step v1.5 LoRA path

LTX-2 conditioning examples are split by conditioning style:

  • ltxvideo2-19b-t2v.peft-lora+first-frame-conditioning shows the shorthand ltx2_* probability fields.
  • ltxvideo2-19b-t2v.peft-lora+intrinsic-conditioning shows the explicit ltx2_intrinsic_conditioning object list.
  • ltxvideo2-19b-t2v.peft-lora+reference-conditioning shows IC-LoRA reference conditioning with coordinate scale overrides.
  • ltxvideo2-2.3-dev-720p-single-gpu.peft-lora+ramtorch adapts the LTX-2.3 720p profile for one GPU with RamTorch transformer-block streaming.

Z-Image conditioning examples:

  • z-image-turbo.peft-lora+canny-conditioning auto-generates Canny edge conditioning data and validates with those references through the IC-LoRA conditioning path.

Cosmos3 examples:

  • cosmos3-image.lycoris-lokr uses RareConcepts/Domokun.
  • cosmos3-edge-image-24g.lycoris-lokr and cosmos3-edge-image-32g.lycoris-lokr target 1024px ARB image LoKr on 24GB and 32GB cards.
  • cosmos3-nano-image-24g.lycoris-lokr, cosmos3-nano-image-32g.lycoris-lokr, cosmos3-nano-image-48g.lycoris-lokr, and cosmos3-nano-image-80g.lycoris-lokr target 1024px ARB image LoKr on progressively larger GPUs.
  • cosmos3-super-image-48g.lycoris-lokr and cosmos3-super-image-80g.lycoris-lokr target 1024px ARB image LoKr for the Super text-to-image flavour.
  • cosmos3-image-48g.lycoris-lokr and cosmos3-image-80g.lycoris-lokr are legacy aliases for Nano 48GB and 80GB profiles.
  • cosmos3-video.lycoris-lokr uses sayakpaul/video-dataset-disney-organized.
  • cosmos3-video-audio.lycoris-lokr uses local synchronized drumming files with audio.auto_split.
  • cosmos3-super-i2v.lycoris-lokr uses nvidia/Cosmos3-Super-Image2Video with video.is_i2v.

Large multi-GPU video examples are split from the standard 24G examples:

  • wan2.1-t2v-14b-480p-8xh100.peft-lora+cp-fa3
  • wan2.1-i2v-14b-480p-8xh100.peft-lora+cp-fa3
  • wan2.1-i2v-14b-720p-8xh100.peft-lora+cp-fa3
  • ltxvideo2-2.3-dev-720p-8xh100.peft-lora+cp-fa3
  • ltxvideo2-2.3-dev-1080p-8xh100.peft-lora+cp-fa3

These profiles assume 8xH100-class hardware, BF16 weights, context_parallel_size=2, and the Hugging Face FlashAttention 3 varlen backend. On A100-class systems, copy the example and change attention_mechanism to flash-attn-varlen-hub before benchmarking.

Modifying and extending an example

You'll want to copy the folder from simpletuner/examples to config before modifying anything, otherwise your changes will conflict with newer example config updates.

cp -R simpletuner/examples/kontext.peft-lora config/kontext.peft-lora

Inside the file config/kontext.peft-lora/config.json you will need to update the locations of output_dir and dataloader_config