How to Train a Custom Style Model in Getimg.ai
A custom model in Getimg.ai is a fine-tune you train on 10–20 of your own images so every future generation inherits that look — same palette, same lighting, same grain. Training runs on their GPUs, costs a flat batch of credits, and takes roughly 15–30 minutes per model.
Why bother training instead of prompting harder?
Because prompts leak. Ask for “soft beige product photography” on ten different days and you get ten interpretations. A fine-tune pins the interpretation down.
Three situations justify the credits:
- A brand with a defined visual identity shipping 50+ images a quarter.
- A recurring character or mascot that must look identical across frames.
- A product line where surface material and lighting must match exactly between shots.
If you need 20 images for one campaign, skip training. Prompting is cheaper. Training pays off somewhere past the second or third reuse.
What dataset actually works?
Ten to twenty images. Fewer than ten and the model overfits; more than twenty adds time without adding much accuracy, and half of them are usually redundant anyway.
Quality rules that matter more than count:
- Same subject or same style, not both. Mixing a character and a background aesthetic produces a muddled model that does neither well.
- Square crops, 1024px minimum. The trainer resizes for you, but resizing a badly framed photo preserves the bad framing.
- Kill duplicates. Five nearly identical shots teach the model nothing new and bias it toward that angle.
- Match output resolution. If final images will be 1024×1024, train on 1024×1024 crops.
- No watermarks, no stock imagery. The model learns the watermark as style, and you cannot fully prompt it away later.
Caption each image with a short sentence. Captions can be as plain as a matte black ceramic mug on an oak table. Generic captions like image 1 work but produce a model that resists direction later.
How do you train the model step by step?
Go to Custom Models in the left sidebar, then Create Model.
- Name it something that survives six months of forgetfulness.
brand-mug-v3-oct2026beatstest2. - Pick a base: SDXL for photoreal texture, SD 1.5 for stylized illustration and lower credit cost per training run.
- Upload the dataset. The uploader accepts JPG and PNG, caps files around 10MB each.
- Set steps. The default 1,000 is a reasonable start. Under 500 you get an undertrained model that ignores the style; over 2,000 you risk baking in the exact poses from your dataset.
- Start training and watch the sample grid it renders during the run. Samples appear every few hundred steps.
One material limitation: you cannot pause or resume a training run, and you cannot edit a finished model. Fixing a bad dataset means training a new model and paying again. Get it right before you start.
How do you test a new model without burning credits?
Use a fixed test prompt. The point is comparison, not beauty.
Something like ceramic object on wooden surface, soft window light, product shot. Run it at 1 credit per image, batch of four, on three models: your new one, the base model, and whatever you used last quarter.
Look for:
- Does the palette match your reference? Open the reference image side by side.
- Does composition change with the prompt, or does every output look like your dataset?
- At strength 0.3 in Image to Image, does the style transfer cleanly onto a new subject?
If every output looks like a copy of training image #3, you overfit. Retrain with fewer steps or a more varied dataset. There is no slider to fix it after the fact.
How do you use the trained model day to day?
Custom models appear in the model dropdown next to SDXL and Flux. Select yours and prompt normally.
Insert the trigger token in your prompt — the unique word you set when creating the model, usually the model name. It biases generation toward the fine-tune. Without it, results drift toward the base model and you have spent credits for nothing.
The two-step workflow that works:
- Generate a batch of four at low resolution with your custom model.
- Take the winner into Creative Upscale at 2x, which re-renders detail at about 4x the credit cost of a base generation.
This keeps exploration cheap. Most of your credits should go to the four variants you throw away, not to the one you keep.
What does it cost, and where does it break down?
Training a custom model costs a fixed credit batch — on the $12 Basic plan (3,000 credits) a single model train plus testing usually eats 5–8% of your month. The $36 Standard plan adds more concurrent jobs, which matters if you iterate fast.
Where it breaks: hands, small text, and anything requiring exact geometric accuracy. A fine-tune cannot fix model-level weaknesses. If your product has a printed label with six words, the fine-tune will still misspell it.
Also worth knowing: models trained on one product line do not generalize. A mug model will produce mugs. Ask it for a chair and you get a chair-shaped mug.
Who should skip custom models entirely?
Solo creators making fewer than 30 images a month. One-off campaign work. Anyone without an existing reference library to train on — you cannot train a brand style before the brand has visual assets, and a model trained on stock imagery is just an expensive stock generator.
Quick Summary
- Custom models need 10–20 square images at 1024px, no duplicates, no watermarks.
- Training takes 15–30 minutes and a fixed credit batch, on Getimg.ai servers.
- Use a trigger token in every prompt or results drift to the base model.
- Under 500 steps undertrains; over 2,000 overfits to dataset poses.
- Training cannot be paused or undone — a bad dataset means retraining and repaying.
- Best fit: brands with 50+ recurring images. Poor fit: one-off projects and solo creators under 30 images/month.
