Kiln

Lightroom for AI Training
Train custom AI image models without a PhD.

Why Kiln? • The Vision • Status

The Problem

Training AI on your own images today looks like this:

RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call

Or this:

network_args:
  rank: 32
  alpha: 16
  conv_rank: 16
  conv_alpha: 8
optimizer_type: prodigy
optimizer_args:
  betas: [0.9, 0.99]
  weight_decay: 0.01
  decouple: true
  use_bias_correction: true
  safeguard_warmup: true
  d_coef: 2

We think that's insane.

You shouldn't need to understand CUDA kernels, learning rates, or LoRA ranks to teach AI what your cat looks like.

Why Kiln?

For everyone:

• Drop photos in. Click train. Get a model that knows your subject.
• Plain English errors. No cryptic ML jargon. Ever.
• Desktop app that feels like professional software, not a research tool.

For power users:

• Full control when you want it. Zero config when you don't.
• Extend everything. Custom modules for datasets, training, inference.
• CLI for automation. GUI for exploration. Your choice.

The Vision

"ComfyUI's power, with actual software design."

Kiln is built around 5 zones—Datasets, Training, Experiments, Generate, Models—each infinitely extensible through modules. The bundled modules handle 90% of use cases. The module system handles the other 90%.

We're starting with one model, done right: Z-Image training that just works.

Then we're opening the floodgates.

Status

🚧 Under Construction

We're building in public. The specification is complete. Implementation is starting.

Want to follow along? Star the repo. Want to help? Check back soon for contribution guidelines.

AI image training is gatekept by complexity.
We're changing that.