LoRA Training AI Generator
Generated on PixelDojo with Flux 2 LoRA Trainer. Produced by PixelDojo's generation pipeline.
To train a LoRA on your own photos you upload ten to twenty of them, write one caption that describes your subject, pick a step count, and start the run. On PixelDojo you do that on a hosted trainer, so there is no GPU to rent and no training script to configure, and the price is fixed per thousand steps: 45 credits on Z-Image, 90 on WAN 2.2 Image, and 160 on Flux 2. The default runs are 2,000 steps on Z-Image for 90 credits, 1,000 steps on WAN 2.2 Image for 90 credits, and 1,000 steps on Flux 2 for 160 credits. At the $0.0625 a credit that every plan charges, that is about $5.63, $5.63 and $10.00 respectively.
Four things that decide whether the run works
Captions decide what you can prompt later
The default caption you write at training time is the phrase that summons the LoRA afterwards. Use something you would never type by accident. A zip upload can also carry a .txt caption per image when different photos need different descriptions.
More steps is the most common mistake
The trainers ship sensible defaults, 2,000 steps on Z-Image and 1,000 on the other two, and pushing well past them tends to overfit. An overfit LoRA reproduces its training photos and stops responding to the rest of your prompt, which is the opposite of what you trained it for.
Person mode and style mode are different jobs
The WAN 2.2 Image trainer separates them on purpose. Person training runs face detection and masking so the run concentrates on the face. Style training switches that off so the run learns brushwork, palette and composition instead. Picking the wrong one wastes the whole run.
Price the experiment, not just the result
Your first LoRA on a new photo set is a test of the photo set. Run it on the 45 credit trainer, look at what came out, fix the set, and only move up to the 160 credit trainer once the photos are proven. That order costs less than being right on the first try elsewhere.
Every price, step count and photo-count guideline on this page was read out of the live trainer configuration on August 26, 2026, the same day the Flux 2 trainer was turned back on after a smoke training run passed end to end in 449.8 seconds.
Why Choose Pixel Dojo for LoRA Training
Professional-quality results with cutting-edge AI technology
Z-Image trainer, 45 credits per 1,000 steps
The cheapest and the quickest of the three. It produces compact rank-16 LoRA files and most runs finish in well under an hour, so photos uploaded in the morning are generating by lunch. The default run is 2,000 steps, which comes to 90 credits, and the step range accepted goes from 10 up to 40,000.
WAN 2.2 Image trainer, 90 credits per 1,000 steps
This one has two modes. Person training turns on face detection and masking, which is what you want when the subject is a face. Style training turns that off so the run learns a visual look instead. Ten to twenty reference images is the guidance, the default is 1,000 steps for 90 credits, and the accepted range is 500 to 3,000.
Flux 2 trainer, 160 credits per 1,000 steps
The most expensive rung and the one to pick for premium character and product LoRAs, because the Flux 2 base renders text crisply and holds fine facial detail. The default is 1,000 steps for 160 credits, and the accepted range runs from 100 to 10,000 steps.
Training runs on our own hardware
LoRA training here is self-hosted rather than brokered out, which is why the pricing is a flat rate per thousand steps instead of a variable compute bill. Your finished LoRA lands in your library and is usable directly in generation, with no weights to upload anywhere else.
How It Works
Collect ten to twenty photos
Ten or more good photos is the working minimum and ten to twenty is the guidance on the WAN 2.2 Image trainer. Vary the angle, the lighting and the background, and keep the subject consistent. A set of twenty near-identical photos teaches less than a set of twelve varied ones.
Pick the trainer that matches your base model
Train on the model you intend to generate with, because a LoRA is tied to its base. Z-Image at 45 credits per thousand steps for speed and cost, WAN 2.2 Image at 90 for a face with masking or a distinct style, Flux 2 at 160 for the sharpest character and product detail.
Choose how the photos go in
Three ways. Upload a zip, which also accepts an optional .txt caption file per image. Upload up to thirty loose images and the zip gets built for you. Or bring a single photo to Character Studio, generate a full training set from it, pick the best shots, and send those back to the trainer.
Write the caption and set the steps
You give the run a display name and one default caption in the form of a photo of your subject in a given style. That caption is what you will prompt with later, so choose a word you would not otherwise type. Then set the step count. The defaults are deliberate, and higher is not better: too many steps overfits and the LoRA stops taking direction.
Start the run, then generate with it
Credits are charged when the run starts and the job is tracked until it finishes. Watch progress under My Trainings. When it lands, the LoRA appears in your library and is selectable in the generation tools straight away.
Loved by creators on PixelDojo
Real feedback from people using PixelDojo, pulled from our in-product surveys.
Because it is awesome
exceptional quality and great overall design of platform and interface. very intuative. love the creative freedom.
Qwen image 2 is amazing!!
Creative freedom, range of tools and options.
I love the training feature
the quality is the best
Explore more AI tools on PixelDojo
AI Tools
Compare & Switch
- Best AI Image Generators
- Best AI Video Generators
- Midjourney Alternatives
- Civitai Alternatives
- Runway Alternatives
- Leonardo Alternatives
- Pika Alternatives
- Luma Alternatives
- Magnific Alternatives
- Veo Alternatives
- Flux Alternatives
- Freepik Alternatives
- Seedance Alternatives
- Seedream Alternatives
- Pixverse Alternatives
- GPT Image Alternatives
- Synthesia Alternatives
- Playground Alternatives
- NightCafe Alternatives
- Canva AI Alternatives
- ElevenLabs Alternatives
- ComfyUI Alternatives
- Fal Alternatives
- Replicate Alternatives
Common Questions
Everything you need to know about LoRA Training
How many photos does a LoRA actually need?
Ten or more high-quality photos works well, and ten to twenty is the guidance on the WAN 2.2 Image trainer. The loose-image uploader accepts up to thirty. Variety matters more than volume: different angles, lighting and backgrounds teach the model what is essential about the subject and what is just the room it was standing in.
What does it cost to train one LoRA?
It is a flat rate per thousand steps. Z-Image is 45 credits, so a 500-step run is about 23 and the 2,000-step default is 90. WAN 2.2 Image is 90 credits, so 500 steps is 45 and 2,000 is 180. Flux 2 is 160 credits, so 500 steps is 80 and 2,000 is 320. At $0.0625 a credit, the Flux 2 default run is $10.00.
Which trainer should I pick for a face?
WAN 2.2 Image in person mode if you want face detection and masking during training, or Flux 2 if you want the finest facial detail and are willing to pay 160 credits per thousand steps for it. Z-Image is the one to start on when you are still learning what your photo set needs, because a mistake there costs 45 credits per thousand steps instead of 160.
How long does training take?
Z-Image runs mostly finish in well under an hour. On the day we turned the Flux 2 trainer back on, our end-to-end smoke training completed in 449.8 seconds and produced a valid LoRA file. Longer step counts take proportionally longer, which is the other reason not to reach for a step count you do not need.
Can I train from a single photo?
Not directly, but there is a path. Take one photo into Character Studio, generate a consistent set of shots of that same subject, choose the ones that hold the likeness, and send those into the trainer as your training set. You are still training on a real set, it is just a set you generated.
What happens to the LoRA after training?
It lands in your library and becomes selectable in the generation tools immediately. You prompt it using the caption phrase you set at training time. Nothing needs exporting, hosting or wiring into an endpoint, and the file stays yours.