> ## Documentation Index
> Fetch the complete documentation index at: https://docs.omnichar.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Train a Z-Image LoRA on your own GPU

> Train a Z-Image LoRA locally, keeping Turbo's 8-step speed with a training adapter or using De-Turbo. About 13GB at 512px, measured on an L40S and a Tesla T4.

Z-Image Turbo is step-distilled, so the trainer offers two ways around that. It needs no 4-bit path: bf16 trains in about 13GB at 512px and about 15GB at 1024px. It trains on the Z-Image Turbo checkpoint a generate node already downloaded.

## Choose a base

> [tab] Turbo + training adapter
>
> Fuses a de-distillation adapter into the base for the run and drops it when the LoRA is saved, which keeps the 8-step speed. Put [ostris/zimage\_turbo\_training\_adapter](https://huggingface.co/ostris/zimage_turbo_training_adapter) in `models/loras/`. Any filename containing `adapter` is detected automatically, or point `INLINE_ZIMAGE_TRAIN_ADAPTER` at a specific file.
>
> Keep these runs short. The adapter slows the distillation breakdown but doesn't stop it.

> [tab] De-Turbo
>
> Trains without an adapter and needs no extra download. Use it when you'd rather not manage another file, or when a run is long enough that the adapter's breakdown would start to show.

## VRAM

| Configuration | L40S (48GB) | T4 (16GB) |
| - | - | - |
| 512px, De-Turbo | 13.1GB | 13.4GB |
| 512px, Turbo + adapter | 13.1GB | 13.4GB |
| 1024px, De-Turbo | 14.9GB | out of memory |
| 1024px, Turbo + adapter | 14.9GB | out of memory |

The two modes peak at the same number, because the adapter is fused into the base before training starts. A 16GB card trains at 512px but not 1024px; a 24GB card does both.

## Speed

The 12-step benchmark at 512px takes 85s on an L40S and 285s on a T4. That makes Z-Image quick to iterate on, which suits the advice to keep adapter runs short.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.