Train custom image styles, characters, or brand visuals and generate consistent assets with Ideogram V4.
A short name shown later in Ideogram 4.0.
Optional phrase to reuse in Ideogram prompts.
Used for images without caption files. Include a unique phrase here if you want a trigger word in prompts later.
More steps can capture the dataset more strongly, but cost and training time scale linearly.
Auto chooses the largest common no-upscale crop. Presets crop images without upscaling.
Sign in to spend tokens
Training Dataset ZIP is required
Your trained LoRA weights and usage instructions will appear here
Ideogram V4 LoRA Trainer trains a custom style, character, or brand look from your own images, so Ideogram 4.0 can generate consistent on-brand assets forever after. On Prizmad you upload a ZIP dataset, set the training steps, and get reusable LoRA weights back.
The Ideogram V4 LoRA Trainer is a training tool, not an image generator. You feed it a ZIP file of example images — your product photography style, a recurring brand character, a specific campaign aesthetic — and it trains a LoRA (a lightweight custom model layer) that captures that look. The output is a set of trained weights you then plug into Ideogram 4.0 to generate new images in exactly that style.
You control the training with a few practical knobs: a LoRA name so you can find it later, an optional trigger word to invoke the style in prompts, a default caption applied to images without their own caption files, a training-steps slider from 100 to 40,000 (more steps capture the dataset more strongly, but cost and training time scale linearly), and a resolution setting where Auto picks the largest common no-upscale crop. An advanced learning-rate field exists, but the default is right unless you're deliberately tuning training behavior.
On Prizmad, the trainer uses the same token balance as Ideogram 4.0 itself — plus AI avatars, video, voiceover, captions, and music. Train once, then generate consistent brand assets from any device using your normal token balance, with full commercial rights to everything you produce.
Train on your best-performing ad creatives so every new Ideogram 4.0 generation matches the visual identity your audience already recognizes.
Train on your product photography so new lifestyle scenes, seasonal variants, and listing images keep the same product look shot after shot.
Teach the model a mascot or spokesperson-style character once, then place it in new scenes, formats, and campaigns without visual drift.
Generate dozens of on-style creative variants for A/B testing — same aesthetic, different hooks, layouts, and formats.
Agencies and freelancers can train one LoRA per client brand and switch between them in Ideogram 4.0 by swapping the weights URL.
Learns a visual style, character, or brand look from a ZIP of example images and packages it as reusable LoRA weights for Ideogram 4.0.
Set an optional trigger phrase (like ACMESTYLE) during training, then drop it into Ideogram 4.0 prompts to invoke the trained look on demand.
A default caption covers images without caption files — include a unique phrase here if you want a trigger word available in prompts later.
Training steps range from 100 to 40,000 (default 1,000). More steps capture the dataset more strongly; cost and training time scale linearly.
Auto chooses the largest common no-upscale crop, or pick square, landscape, portrait, widescreen, ultrawide, phone, or banner presets — images are cropped, never upscaled.
A completed run produces a weights URL you paste into Ideogram 4.0's 'Use trained LoRA' option, with an adjustable strength slider at generation time.
Collect your training images — consistent examples of the style, character, or product look you want the model to learn — and pack them into a single ZIP file.
Upload the ZIP as your training dataset in the tool above.
Name the LoRA (for example, 'Acme product style') so it's easy to identify later, and set an optional trigger word like ACMESTYLE to reuse in Ideogram prompts.
Write a default caption for images that don't have their own caption files. If you want a trigger word available in prompts later, include a unique phrase here.
Set training steps. The default of 1,000 is a sensible starting point; more steps capture the dataset more strongly, but cost and training time scale linearly.
Choose a training resolution — Auto picks the largest common no-upscale crop across your dataset — and start the run. Training takes roughly 15–45 minutes.
When training finishes, copy the weights URL from the result, open Ideogram 4.0, enable 'Use trained LoRA', paste the URL, and dial in the LoRA strength to start generating in your custom style.
A training run is listed at 338 tokens, and cost scales linearly with the training steps you choose — the default is 1,000 steps. You pay to train once; the resulting weights are reusable across as many Ideogram 4.0 generations as you like.
The trainer is unlocked on the same token balance as Ideogram 4.0, AI avatars, video, voiceover, and music — no separate Ideogram account, no API key setup, no extra billing.
Training and generating are separate costs: after training, images made with your LoRA in Ideogram 4.0 use that tool's own image pricing. If tokens run out mid-project, a one-off top-up covers it without an upgrade.
No — training consumes tokens from your Prizmad balance. A run is listed at 338 tokens, and cost scales linearly with the number of training steps you set. There is no separate Ideogram account or API key required; it's available for Prizmad tokens.
Yes. Assets you generate on Prizmad with your trained LoRA are yours to use commercially — in paid ads, listings, landing pages, and printed materials.
A completed run returns trained LoRA weights (a .safetensors URL). You paste that URL into Ideogram 4.0's 'Use trained LoRA' setting and adjust the LoRA strength slider to control how strongly your trained style affects each generation.
Roughly 15–45 minutes, depending on your settings. Training runs asynchronously, so you can keep working while it completes. More training steps mean longer runs — training time scales linearly with steps.
The default is 1,000 steps, on a slider from 100 to 40,000. More steps capture the dataset more strongly, but both cost and training time scale linearly — so start at the default and increase only if the trained style comes out too weak.
Consistent example images of the style, character, or product look you want to capture. Images without caption files use your default caption — include a unique phrase there if you want a trigger word to invoke the style in prompts later. The resolution setting crops images without upscaling; Auto picks the largest common crop.
If well-written prompts get you the look you need, Ideogram 4.0 alone is enough. Train a LoRA when you need the same specific style, character, or brand look reproduced consistently across many generations — something prompting alone can't guarantee.
Yes — each training run is independent and produces its own named weights. You can keep separate LoRAs for different brands, products, or campaign styles and switch between them in Ideogram 4.0 by pasting the relevant weights URL.
Published 2026-07-16 · Last updated 2026-07-16