liveBook · Manning

AI-Toolkit

As explained in Local Image Generation

An open-source toolkit for configuring and running training of generative-model adaptations such as Flux LoRAs.

AI-Toolkit is open-source software for training Flux LoRAs: small adaptations that teach a model a custom subject or style without retraining everything. Its browser UI handles the practical loop—organize a dataset, configure a job, run training, and inspect results. Expect a hungry GPU: Flux LoRA training requires 24 GB of VRAM in this workflow. section 13.0 section 13.1

Figure 13.2 The AI-Toolkit as it appears in the browser.

The browser interface is the control room for datasets, training jobs, and training results. section 13.1

A basic run looks like this:

  • Clone and install the project, then run npm run build_and_start from its ui directory. Open the local server address in a browser. section 13.1
  • Prepare high-quality images of the subject across poses and environments, then caption them descriptively. Careful labeling is tedious but central to the result. section 13.6
  • Create a job, give the LoRA a meaningful name and trigger word, select Flux.1, and choose the correct dataset. Most remaining settings can initially stay at their defaults. section 13.3
  • Save checkpoints throughout training and compare sample-prompt images. Prefer the earliest checkpoint that represents the subject correctly before overtraining distorts other subjects. section 13.3 section 13.6

Training a custom dog LoRA

A distinctive dog can make training results easier to judge: use varied photos, captions that describe each image, and sample prompts that test the dog in several scenarios. The winning checkpoint is not automatically the final one; it is the earliest version that captures the subject without unwanted distortions. section 13.0 section 13.6

How Local Image Generation covers it

How deep it goes
Walks through installation, UI startup, job settings, checkpoints, and sample-prompt debugging while leaving most configuration at defaults.
Examples
Uses a Silken Windhound LoRA and browser-based job-creation screens to demonstrate the workflow.
What it assumes
Assumes access to a locally installed project and 24 GB VRAM, while teaching the training workflow from dataset preparation onward.
What you can do after
Operate/configure Flux LoRA training jobs, preserve checkpoints, and choose a usable result from sample images.
How it connects
Connects Flux LoRAs with training data, image captions, trigger words, checkpoints, overtraining, and sample prompts.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Local Image Generation. The section links point to the parts of the book it draws on.