LoRA Fine Tuning AI Generator
Generated on PixelDojo with LoRA trainers. Produced by PixelDojo's generation pipeline.
Six base models on Pixel Dojo accept a custom LoRA you train yourself: Flux, Flux 2, Z-Image, WAN 2.2 Image, Qwen and SDXL. Prices at each trainer's default step count run from 4 credits for SDXL up to 160 credits for Flux 2, and every one of them charges by training step rather than by result. This page lists all six with the exact formula each route uses, so you can price a run before you upload anything. The numbers come straight from the trainer code, not from a marketing sheet.
The six trainers and what each one charges
SDXL
2 credits per 500 training steps. The route defaults to 1000 steps, so a default run is 4 credits. That makes it the cheapest way to find out whether your image set is good enough before you spend real credits on a newer base.
Qwen
A flat 32 credits per training run, no matter how many steps you choose. It is the only trainer here that does not scale with steps, which makes budgeting simple when you want to push the step count high.
Flux
16 credits per 500 steps. The route defaults to 2000 steps, so the standard run is 64 credits. This is the trainer with the longest track record for photo realistic people.
Z-Image
45 credits per 1000 steps, defaulting to 2000 steps for 90 credits. A short 500 step test comes to 23 credits. Steps are accepted anywhere from 10 to 40000, and the trainer offers three focus modes: content, style, or balanced.
WAN 2.2 Image
90 credits per 1000 steps, and the route defaults to 1000 steps, so a default run is 90 credits. The trainer estimates roughly one minute of wait per 100 steps, which puts a default run near ten minutes.
Flux 2
160 credits per 1000 steps, defaulting to 1000 steps. Step counts are clamped to the range 100 to 10000 and rounded to the nearest 100, so a 2000 step run prices at 320 credits. It is the most expensive trainer in the list and the newest base.
Three of them run on our own training hardware
Flux 2, WAN 2.2 Image and Z-Image train on infrastructure we operate ourselves. They came back that way in August 2026 after a pause. The other three trainers were never moved.
What a Z-Image LoRA file looks like
The Z-Image trainer outputs a rank 16 file. That was checked tensor by tensor against the older hosted trainer's output in August 2026, and the layouts matched, which is why trained files load at the same strength either way.
How many images you need
Ten or more good images is the working guidance in the trainer itself. You can upload a zip with an optional .txt caption file next to each image, or hand the trainer up to 30 loose images and let it build the zip for you.
Where a trained LoRA gets used
A finished LoRA shows up in the model picker on the generation pages that support one, including Flux, Flux 2 Flex, Z-Image Turbo, Qwen Image, PonyXL and WAN 2.2 video. Training on the base you plan to generate with is the whole point of picking a trainer.
Six trainers, six different price formulas. The cheapest default run is 4 credits. The most expensive is 160.
Why Choose Pixel Dojo for LoRA Fine Tuning
Professional-quality results with cutting-edge AI technology
Priced per step, not per image
Every trainer except Qwen multiplies a per-step rate by the steps you choose. Halving the steps halves the price.
4 credits buys a real test run
A default SDXL training is 4 credits. That is cheap enough to use as a check on your dataset before committing to a 160 credit Flux 2 run.
You keep the file
The trained LoRA is yours, stored against your account, and selectable from the generation pages that accept one.
How It Works
How to price and start a training run:
Pick the base you will generate on
A LoRA only loads on the model it was trained against. Decide whether your finished images will come out of Flux, Flux 2, Z-Image, WAN 2.2 Image, Qwen or SDXL first, then open that trainer.
Do the step math before you upload
Multiply the trainer's rate by your steps. Z-Image at 45 per 1000 is 23 credits for a 500 step test and 90 for the 2000 step default. Flux 2 at 160 per 1000 is 160 for the default and 320 at 2000 steps.
Send a zip or a loose set
Ten or more images works. A zip can carry a .txt caption beside each file, or you can upload up to 30 images and let the trainer build the zip. Credits are charged when the run starts and refunded if it fails.
Loved by creators on PixelDojo
Real feedback from people using PixelDojo, pulled from our in-product surveys.
es facil de comprender asi uno no sea programador
amazing web site
phenomenal site. would highly recommend
Amazing features, easy to use, privacy
the number of options, and especially the quick response to questions on Discord
Love you guys!!
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 Fine Tuning
Which AI models can be fine tuned with a LoRA on Pixel Dojo?
Six: Flux, Flux 2, Z-Image, WAN 2.2 Image, Qwen and SDXL. Each has its own trainer with its own credit formula. Nothing else in the catalog takes a LoRA you trained yourself.
What is the cheapest LoRA to train?
SDXL. It charges 2 credits per 500 steps and defaults to 1000 steps, so a default run costs 4 credits. Qwen is next in practical terms at a flat 32 credits regardless of step count.
How much does a Flux 2 LoRA cost?
160 credits per 1000 training steps. The trainer defaults to 1000 steps, so the standard run is 160 credits and a 2000 step run is 320. Step counts are rounded to the nearest 100 and clamped between 100 and 10000.
How many photos does a LoRA need?
Ten or more high quality images is the guidance in the trainer. Upload them as a zip with optional caption text files, or send up to 30 loose images and the trainer packages them for you.
Can I use a trained LoRA in the public API?
Yes. Trained files are attached to your account and selectable on the tools that accept a LoRA, and those tools are callable with the same API key and the same request shape as every other model in the catalog.
Which base should I pick for a person?
Flux and Flux 2 are the photo realistic bases, and Z-Image is the cheapest of the modern three at 45 credits per 1000 steps. Train on whichever base you plan to generate with, because a LoRA only loads on the model it was trained against.