LoRA training
Train a LoRA on your own GPU, on a canvas
Inline Studio has a built-in LoRA trainer for Z-Image and Krea 2. Turn a folder of images into a trained model without scripts or notebooks, on your own GPU. Free and open source. A Krea 2 LoRA fits a 16GB card at 512px with the 4-bit base.
How to train a LoRA, step by step
- 1
Build a dataset
On the Trainer tab, create a dataset, give it a trigger word, and drag in 20 to 50 images. A single subject wants fewer; a style wants more.
- 2
Auto-caption locally
Caption the set with the local captioner, or leave captions off and lean on the trigger word. The trigger is prepended to every caption during training.
- 3
Pick a model and base
Choose Z-Image or Krea 2, then a base. Krea 2 RAW is the recommended path: train on RAW, generate with Krea 2 Turbo, and the same LoRA applies to both.
- 4
Wire the graph and start
Connect Load Dataset into Caption into Train LoRA, set rank and steps, and hit Start. The Graph node plots loss live and a Resources monitor tracks VRAM.
- 5
Generate with your LoRA
The finished LoRA lands in models/loras, ready to drop onto the generation canvas with a LoRA loader node.
Z-Image and Krea 2
Krea 2 (recommended)
Train on the undistilled RAW base, then generate with Krea 2 Turbo. The LoRA carries over unchanged. A 4-bit base makes the 26GB model fit a 16GB card at 512px, about 12GB peak. If you only hold Turbo, add the Krea 2 training adapter.
Z-Image Turbo
Keep the 8-step speed with the Turbo training adapter, or switch to De-Turbo with nothing extra to download. Trains at 512px in about 13GB, so it fits a modest card.
LoRA training VRAM, measured
What fits on which card, so you know before you start.
| Model and resolution | Fits 16GB? | Peak VRAM |
|---|---|---|
| Z-Image, 512px | Yes | ~13GB |
| Z-Image, 1024px | No, needs 24GB | ~15GB |
| Krea 2 (4-bit base), 512px | Yes | ~12GB |
| Krea 2 (4-bit base), 1024px | No, needs 32GB | ~28GB |
Peak allocation on a Tesla T4 (16GB) and L40S, rank 16, batch 1, gradient checkpointing on. 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.
- Base precision (Krea 2)
- 4-bit for a 16GB card (about 12GB peak), or full precision on 32GB and up. Auto picks for you.
- 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
Which base models can I train a LoRA for?
Z-Image Turbo and Krea 2. For Krea 2, train on the undistilled RAW base and generate with Krea 2 Turbo, the same LoRA carries over unchanged. For Z-Image, train with the Turbo training adapter to keep its 8-step speed, or in De-Turbo mode with nothing extra to download.
Can I train a Krea 2 LoRA on a 16GB GPU?
Yes, at 512px with the 4-bit base. It peaks around 12GB, measured on a Tesla T4, so it fits a 16GB card. Krea 2 at 1024px needs roughly 32GB, so train at 512. A LoRA trained at 512 still applies at any generation resolution.
How much VRAM does LoRA training need?
Z-Image trains at 512px in about 13GB and at 1024px in about 15GB. Krea 2 with a 4-bit base trains at 512px in about 12GB and at 1024px in about 28GB. These are measured peak allocations on a Tesla T4 and L40S at rank 16.
Do I need to write captions?
No. You can auto-caption the dataset locally, or turn captions off and rely on the trigger word alone, which works well for a single subject. The trigger word is prepended to every caption during training.
Can I stop and resume a training run?
Yes. Stopping flushes a checkpoint with the adapter weights, optimizer and RNG state, and step count, so resuming continues from the exact step you paused. Runs interrupted by a crash recover on their own.
Train your first LoRA
Free and open source. Runs on macOS, Windows, and Linux.