Most weak LoRAs are not the trainer's fault. They come from ten near-identical selfies, a blurry product shot or a base model you never generate with. This guide covers the parts you control when you train a LoRA online: the dataset, the trigger word, the base model and the steps, plus what it costs and when a local, no-cost setup is the better choice.
Quick answer: Collect 15–30 sharp, varied images of one subject, pick the base model you will generate with, set one trigger word, keep the default steps for a first run and train in the browser. On Imagera's LoRA Trainer the credit cost of the run is shown before you start, and the finished model is ready to generate with in the same account.
1.Can you train a LoRA online for free?
Not on Imagera: training runs on paid cloud GPUs and is billed in credits. The no-cost route is local training with an open-source trainer such as AI Toolkit (MIT licence), which supports Krea 2, FLUX, Qwen-Image, Z-Image, Wan 2.1/2.2 and MiniMax H3 among others. The catch is hardware and setup: its example configurations are written for 24 GB graphics cards, and you manage drivers, Python environments and dataset formatting yourself.
A browser trainer makes sense when you do not own that card, or when the hours of setup cost more than the run.
2.Step 1: Build the dataset
The dataset decides the result more than any setting.
For a person or character (15–30 photos):
- Mix close-ups, head-and-shoulders and a few full-body shots.
- Vary angle, expression, lighting and background.
- One person per photo. No sunglasses, heavy filters or watermarks.
For a product (10–20 photos):
- Front, back, sides and three-quarter views of the same variant.
- Real photos, sharp and well lit; clean backgrounds help.
For a style (20–50 images):
- The same look applied to different subjects, so the model learns the look, not one scene.
- Keep it consistent: a mixed set trains a mixed style.
Use images of at least 768–1024 px on the short side. Do not upscale small images to hit that number; it adds noise, not detail.
3.Step 2: Choose the base model before you train
A LoRA only works on the model it was trained against, so start from how you will generate:
| You want | Trainer on Imagera | Dataset guidance shown in the trainer |
|---|---|---|
| Photoreal people and products | Krea 2 Turbo | 15–30 sharp photos of one subject or style |
| Fast, sharp all-rounder | Z-Image Turbo | 15–25 clean photos of one subject |
| Faces, products and styles | FLUX.1 Dev | 15–30 varied photos of your subject or style |
| A consistent illustration or brand style | FLUX.2 Klein 9B | 10–50 images in one consistent style |
| Posters, logos, lettering | Ideogram 4.0 | 10–30 images of one style, subject or brand |
| Teach an edit (before → after) | Qwen Edit 2509 | Before/after image pairs |
| A person in motion | LTX-2.3 Character | Clips or photos (not mixed) |
| Cinematic motion or style in video | WAN 2.2 Video | 10–20 short clips and/or photos |
Video trainers, including the four MiniMax H3 modes, are covered in MiniMax H3 LoRA training.
4.Step 3: Trigger word and captions
Pick one short, unusual trigger word (for example ohwx person or brandlook style) and use it every time you prompt. Captions tell the trainer what is not the subject: if the background changes between photos, describing it keeps the background out of the LoRA. Some trainers, such as Krea 2 Turbo, offer automatic captioning in Object/Character, Style or Custom modes.
5.Step 4: Steps and resolution
- Steps: every trainer on Imagera has a 1,000-step minimum, which is the right first run for most image datasets. Raise steps only when the subject is recognisable but soft. If every output looks like a copy of one training photo, you overtrained: retrain with fewer steps or more varied images.
- Resolution: trainers that offer it (Krea 2 Turbo trains at 768 or 1024 px) give sharper detail at the higher setting, for a higher credit cost.
6.Step 5: Train and test
Start the run from the trainer or Sandbox. Image trainers typically finish in about 8 to 20 minutes, depending on the model. Then test with prompts that change everything except the subject:
ohwx person hiking a mountain ridge at sunrise, wide shotohwx person in a studio portrait, black background, soft key lightohwx person as a watercolour illustration
| Problem | Likely cause | Fix |
|---|---|---|
| Doesn't look like the subject | Too few or too similar images | Add varied photos, retrain |
| Every image looks like one training photo | Overtrained | Fewer steps or more variety |
| Only works in one pose | Dataset lacks variety | Add angles and poses |
| Background or clothing always the same | Captions didn't describe them | Caption what changes |
| Style bleeds into everything | Strength too high | Lower strength to about 0.6–0.8 |
Generate with the trained model in LoRA Generate.
7.How much does LoRA training cost?
On Imagera the trainer shows the exact credit cost before the run; it scales with the trainer, the number of steps and options such as training resolution and auto-captioning. Plans start at $19.99/mo for 400 credits a month on the Starter plan, and credit packs start at $39.99 for 350 credits (pricing).
| Route | Money | Setup | Hardware |
|---|---|---|---|
| Local open-source trainer | No software cost | Hours: drivers, Python, configs | Your own GPU (example configs target 24 GB) |
| Rented cloud GPU | Hourly rental | You run the trainer yourself | Rented |
| Imagera LoRA Trainer | Credits, quoted before the run | Upload and start | None |
New to adapters? Start with what a LoRA is and what it can learn.
