How Marketing Teams Train a Custom Image Model Without Standing Up a GPU Box

General image models leave brand teams in a familiar place. You can get a beautiful picture that does not quite look like your product, your store, or the person who is supposed to appear in every campaign. The usual fix is to hire more retouching, or to stand up your own training box, install a stack, and hope someone on the team wants to mind it.

A lighter option is to train a LoRA in the browser. A LoRA is a small adapter that steers a base model toward a look, a product or a character. The idea itself is not new. The original LoRA paper, "Low-Rank Adaptation of Large Language Models", described a way to adapt a large model with a modest number of extra weights. Image tools later picked up the same approach. The change is that you no longer have to run the training job on a machine you own.

PixelDojo's LoRA page walks through that hosted path. Creators upload an image dataset and run training sessions on the platform, which supplies the compute. You can start from a single image, a set of images, or a zip file dataset. From there you adjust training layers, set the LoRA rank, fine-tune other parameters, watch the job, and export the custom model when it finishes. Trained LoRA files can be applied to base models inside PixelDojo's generation tools, and you can manage more than one dataset from a dashboard.

The trainer on that page is called the FLUX Model Trainer. A free account is enough to open it. The page also says the only local requirements are a stable internet connection and a modern web browser.

"Free" and credits are not the same claim

Read the pricing language carefully. The marketing copy and the FAQ do not say quite the same thing. Some sections of the page talk about starting at no cost. The FAQ is more precise: LoRA training uses credits, and you can train as many custom models as your monthly credits allow. PixelDojo's pricing page lists LoRA training among the tools included with every plan, and plans start at $10 a month for 160 credits. Pro is $25 for 500 credits, Pro Max is $50 for 1,100, and credits roll over while you stay subscribed. The LoRA page does not publish a per-job credit price, so look at the cost shown before you start a run rather than assuming a training job is free.

A free PixelDojo account still gets three photo or video effects a day, with a small watermark. That offer is about effects, not about unlimited training.

The page says models can be trained in as little as 7 minutes, and that times can vary. Treat the 7 minute figure as the vendor's best case, not a guarantee, especially if your dataset is large or you raise the rank.

The dataset is the real work

The training interface will not save a messy brief. If you want a product LoRA, photograph the product consistently: the same object, clean backgrounds, more than one angle, no extra logos fighting for attention. If you want a style LoRA, keep the style consistent and leave out images that belong to a different look. If you want a person to appear in later campaigns, you need more than a single flattering headshot, and you need permission.

That last point is not optional. PixelDojo's terms, version 6, effective 4 November 2026, say that when you train a custom model you promise you hold the rights and consents needed for every item in the dataset, including consent from every real person shown. The trained model is treated as output under those terms. Tools that work from a real person's face, voice or likeness require that person's consent. Do not train on a customer's photos, a colleague's face, or a scraped set of brand images unless you already have the rights.

The terms also assign you whatever rights PixelDojo holds in the output. They do not promise that the result is unique or free of other people's rights, and you are responsible for how you use it. Where a platform or the law requires you to label synthetic media you publish, you must label it. The LoRA page's FAQ says trained models can be used in personal and commercial projects, but the contract is the document to read before a client campaign.

A workflow a marketing team can actually run

Keep the first experiment small:

  1. Pick one job, such as "this bottle, on a clean background, in our usual light," not "our entire brand system."
  2. Gather a tight dataset you already have the rights to.
  3. Start with a modest LoRA rank, run one job, and generate a handful of test images in the same studio.
  4. If the object or the face drifts, adjust the dataset before you raise the rank.
  5. Export the LoRA you want to keep, and note which base model you trained it for.

The page also says you can apply the trained file to base models inside the same generation tools, so you do not have to leave the studio to see whether the LoRA actually steers the look. If the first batch of test images drifts, fix the dataset before you train again. A LoRA is a small adapter, not a replacement for a brand style guide. It will not invent consistency that was never in the photos.

You can train a LoRA online with PixelDojo without buying a GPU workstation or learning a local trainer. Speed is not the useful test. What matters is whether the next ten images still look like your product, and whether you would be comfortable putting your name under the dataset you used.