> ## 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 Krea 2 LoRA on a 16GB GPU

> Train a Krea 2 LoRA locally on the undistilled RAW base. A 4-bit base fits the 26GB model into about 12GB at 512px, measured on a Tesla T4.

Krea 2 is the model to pick if you have the choice. You train on the undistilled RAW base, so there's nothing to fuse and nothing to drift, and the LoRA carries over to Krea 2 Turbo unchanged. A 4-bit frozen base brings the 26GB model down to about 12GB at 512px, so a 16GB card can train it.

## Choose a base

> [tab] Krea 2 RAW (recommended)
>
> Put `krea2_raw_bf16.safetensors` in `models/diffusion_models/`, train on it, then generate with the **Krea 2 Turbo** node. The published [skin LoRA](https://huggingface.co/inlineresearch/skin-lora-krea-2-raw) was trained this way.

> [tab] Turbo + training adapter
>
> For when you only have Turbo. Put [ostris/krea2\_turbo\_training\_adapter](https://huggingface.co/ostris/krea2_turbo_training_adapter) in `models/loras/`, or point `INLINE_KREA2_TRAIN_ADAPTER` at it. The adapter is fused into the base before training starts, so it peaks at the same VRAM as RAW.

## VRAM

| Configuration | L40S (48GB) | T4 (16GB) |
| - | - | - |
| 512px, 4-bit base | 11.7GB | 11.9GB |
| 512px, bf16 base | 30.4GB | out of memory |
| 1024px, 4-bit base | 27.8GB | out of memory |
| 1024px, bf16 base | out of memory | out of memory |

Leave **Base precision** on Auto: it sizes the base plus its activations at your resolution against your card. 1024px needs about 32GB, and no setting closes that gap, because activations scale with image tokens and gradient checkpointing and memory-efficient attention are already on. Train at 512 on a small card. The LoRA still applies at any generation resolution.

System RAM stays low. Checkpoints are read tensor by tensor, so Krea 2 trains in about 3GB of host RAM.

## Speed

* The 12-step benchmark at 512px with a 4-bit base takes 192s on RAW and 219s through the Turbo adapter on an L40S, against 824s and 872s on a T4.
* A 1500-step run is roughly 40 minutes on an L40S and about 3 hours on a T4. Turing has no native bf16, so a T4 runs the same work about four times slower.

## Settings that worked

The [skin LoRA](https://huggingface.co/inlineresearch/skin-lora-krea-2-raw) trained on Krea 2 RAW at rank 16 / alpha 16, full scope, 1500 steps, batch 1, learning rate 1e-4, at 1024px, with caption dropout 0.05. Its 26-pair dataset is [published](https://huggingface.co/datasets/inlineresearch/krea2-skin-lora). It generates on Krea 2 Turbo at 28 steps, guidance 4.5.

## Licence

Krea 2 is under the [Krea AI Community License](https://www.krea.ai/krea-2-licensing). Read it before shipping anything commercial.


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