Train a LoRA locally, on your own GPU

How to train a LoRA, step by step

1. 1

   Build a dataset

   On the Trainer tab, create a dataset, give it a trigger word, and drag your images in. The skin LoRA used 26 close-up and half-face photos chosen for real, unretouched skin across tones, ages, and lighting. 20 to 50 is a good range: a single character wants fewer, a broad style wants more.

   

   

   

   Download the dataset from Hugging Face
2. 2

   Caption the set

   Every image gets a caption that starts with the trigger, then a shared phrase, then a plain description of the frame. Auto-caption locally or write them yourself. The trigger is prepended to each caption during training, so it becomes the switch that turns the style on at generation time.

   image 1: inline-skin-lora, detailed skin texture, half-face close-up, bright red hair, very fair skin densely covered in freckles, blue-grey eye, pale lashes, cool tone, natural realistic texture, teal background

   image 2: inline-skin-lora, detailed skin texture, close-up of a young woman looking upward, warm tan skin, smooth dewy complexion, glossy highlight on the nose, brown hair, natural lips, subtle fine texture, soft flattering light

   image 3: inline-skin-lora, detailed skin texture, moody low-key half-face portrait, fair skin, blue eye, curly hair, glossy highlights on the nose and forehead, natural texture, deep shadow falling across the face, dark background

3. 3

   Pick a model and base

   Choose the architecture first (FLUX.2, Z-Image, Krea 2, MiniMax H3 or LTX-2.5), then a base within it. Every one trains on an undistilled base, because training on a step-distilled model breaks the distillation down. The skin LoRA trained on Krea 2 RAW and generates on Krea 2 Turbo. Leave base precision on Auto; the settings table below says what it picks.

4. 4

   Set the config and start

   Wire the three nodes, set the config, and press Start. The skin LoRA used rank 16 at full scope (attention and feed-forward), alpha 16, 1500 steps, batch size 1, learning rate 1e-4, at 1024px, with caption dropout 0.05, no flip, and a checkpoint every 250 steps. The Graph node plots the loss live, a Resources monitor tracks VRAM, and a stopped run flushes a checkpoint so you can resume.

   [ Load Dataset ] --> [ Caption ] --> [ Train LoRA ] --> [ Graph ]
                                             |
                                             +--> Resources (VRAM monitor)

5. 5

   Generate with your LoRA

   The finished LoRA lands in your models/loras folder, ready to drop onto the generation canvas with a LoRA loader node. Prompt with the trigger phrase and generate on Krea 2 Turbo at 28 steps, guidance 4.5. To try this exact one first, download it from Hugging Face and drop the file into models/loras.

   Download the skin LoRA from Hugging Face

FLUX.2, Z-Image, Krea 2, MiniMax H3 and LTX-2.5

Pick the architecture in the Trainer's Adjust panel, then a base within it. Each one trains on an undistilled base and the LoRA carries over to the fast checkpoint unchanged.

A real example: the skin LoRA

Our first published LoRA was trained here on Krea 2 RAW from 26 images. It brings back real pores, fine hair, and natural texture across skin tones, ages, and lighting. Both the LoRA and the dataset it trained on are free to download.

Click to zoom

The recipe

The exact settings behind these renders.

Base model

Krea 2 RAW (bf16), carries over to Turbo

Base precision

Auto (4-bit on 16GB, full precision on 32GB+)

CPU offload

Auto

Method

LoRA (PEFT)

Rank / alpha

16 / 16

Scope

Full (attention + feed-forward)

Steps / batch

1500 / 1

Learning rate

1e-4

Resolution

1024px

Caption dropout

0.05, no flip augmentation

Checkpoint every

250 steps

Dataset

26 image and caption pairs

Trigger

inline-skin-lora, detailed skin texture

Generation

Krea 2 Turbo, 28 steps, guidance 4.5

How it holds up

The same LoRA across six conditions: three in the close-up framing it trained on, three well outside it. Texture holds in all six.

One limit worth prompting around, visible above: framing pulls toward head-and-shoulders unless you ask for something wider. The model card has the full write-up.

LoRA training VRAM, measured

What fits on which card, so you know before you start.

Peak allocation on a Tesla T4 (16GB) and L40S, rank 16, batch 1, gradient checkpointing on. The FLUX.2 rows were measured on the L40S and interpolated to 16GB. MiniMax H3 was measured on an L40S, an L4 and a T4. Every resolution reads the same because the peak is the caption pass rather than the training, and so do clips: a 4.5s clip packs 8,232 rows against a still's 293 and still peaks at 20.4GB. A card too small for the 32B conditioner runs it on the CPU instead, which drops the peak to 12.7GB but costs 16.2s a step and a 19 minute caption pass. LTX-2.5 was measured on the L40S only: its 22B base is 38GB at bf16 with no 4-bit path to fall back on, so it needs a 48GB card and a 24GB one will not do. Leave headroom for your GPU's context. A LoRA trained at 512 still applies at any generation resolution. See the full matrix.

Recommended settings for 512px

Resolution

512 to fit a 16GB card. A LoRA trained at 512 still applies at any generation resolution. FLUX.2 is cheap enough that 1024 also fits, at about 9.9GB.

Base precision (Krea 2)

4-bit for a 16GB card (about 12GB peak), or full precision on 32GB and up. Auto picks for you.

Base precision (FLUX.2)

Leave it on Auto, which stays at bf16. 4-bit gains nothing here and costs about 10 percent a step, because the peak sits in the caption and latent caching pass rather than in the transformer.

Rank and alpha

16 for a single subject, 32 for a style. Keep alpha equal to rank.

Steps

1000 to 1500 is enough for strong likeness, roughly 30 to 40 steps per image.

Learning rate

1e-4 is the standard starting point for a LoRA.

LoRA training FAQ