How to Train AI Image Models in 2026

Scenario setup

You hold 40 product photos of a new sneaker line. The campaign launches in 90 minutes and you need 200 consistent images across 12 angles. Flixly does not offer model training. Instead the platform supplies 50+ ready models that accept reference images for character or style control.

Prepare your reference set

Start at the dashboard. Upload the 40 sneaker photos to a new project folder. Name the folder sneaker-2026-ref. Keep files under 8 MB each and use 1024x1024 resolution for best results with the models listed below.

Select the strongest 12 images that show front, side, top, and detail views. Delete duplicates and any shots with heavy shadows. This curated set becomes the visual anchor for every generation that follows.

Choose the base model

Open the AI Image Generator page. From the model dropdown pick GPT-Image 2.0 first. It handles product geometry well. If the first batch shows color drift, switch to FLUX Kontext for the next round because it locks material textures more tightly.

Set the prompt to "sneaker on white background, exact same model as reference images, studio lighting, 8k detail". Paste the same prompt for every run so only the camera angle changes.

Run the first generation pass

Enter the prompt, attach the 12 reference images, and set batch size to 20. Generation time averages 8 seconds per image on GPT-Image 2.0. Download the batch and open the files in a grid view.

Check for consistency in logo placement and sole tread pattern. Any image that deviates gets deleted. Keep 16 good frames from the first pass.

Refine with image-to-image

Move to the Image to Image tool. Load one approved frame as the source. Lower the denoising strength to 0.35 so the model respects the original composition while allowing small angle adjustments.

Repeat for the remaining 15 frames. Each new output lands in the same project folder and inherits the file-naming convention you set earlier.

Add variation with headshot and logo tools

Switch to the AI Headshots page to generate matching lifestyle shots of a model wearing the sneakers. Use the same reference set so the shoe stays identical. Generate 30 images at 512x768.

Then open Logo Generation and create 10 alternate logos that match the sneaker colorway. Export them as transparent PNGs at 2048 pixels wide.

Composite final assets

Return to the dashboard and use the product-mockup tool (linked from the image tools section) to place the logos onto the generated sneakers. Run a batch of 40 composites. Each composite finishes in 12 seconds.

Verify and export

Open the project gallery. Sort by creation time and scan the last 40 files. Confirm every image contains the correct sole tread and color code. Flag any file that fails and regenerate it with a new seed.

Export the approved set as a zip named sneaker-campaign-2026-final.zip. The archive contains 200 images totaling 1.8 GB.

Step-by-step instructions

1. Create a new project folder and upload your 40 reference photos at 1024x1024.
2. Select the 12 strongest images that cover all key angles.
3. Navigate to the text-to-image tool and choose GPT-Image 2.0.
4. Write a fixed prompt that names the product and lighting style.
5. Attach the 12 references and generate a batch of 20.
6. Review the batch, delete failures, and keep the best 16 frames.
7. Move to image-to-image with denoising at 0.35 for controlled variations.
8. Export the final 200 images in a single zip archive.

Model comparison table

Outcome

You now own a 200-image asset pack that matches the original 40 references exactly. Return to the same project any time you need more angles by repeating the reference attachment step on the AI Image Generator page.