Training a Qwen-Image-2512 Face LoRA in AI Toolkit: 3,000 Steps, Rank 64, Batch 1
The most reusable asset in image generation right now is not a workflow file. It is a checkpoint that already knows one specific face. This episode is the author's on-camera recipe for building exactly that with AI Toolkit and Qwen-Image-2512, from an empty cloud account to a finished LoRA. Evidence note: this episode ships no workflow JSON . Every number below is a value the author states in the…
This article provides a detailed, step-by-step guide on training a Qwen-Image-2512 Face LoRA in AI Toolkit from an empty cloud account to a finished LoRA. The author emphasizes that the most reusable asset in image generation is a checkpoint that already knows a specific face.
The process begins by renting a machine, installing AI Toolkit, and pulling the weights with ModelScope. The base model path is set to the Qwen-Image-2512 directory, and various settings such as LoRA rank, batch size, gradient accumulation, resolution, timestep type, EMA, and float8 are adjusted. The training task is named after the trigger word and points to the 2512 model path.
The author specifies the learning rate, optimizer, and save interval. Sampling is turned off, but it can be set to every 500 steps if desired. The article concludes by emphasizing that the numbers used in this recipe are specific to training a single face on one A800 and that the defaults are not recommended for other use cases.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
