Product image generation with local AI: without data leaving the company

Mehmet Akar 4 min read
Examples of textile product visuals generated with local AI

Generating product images with local AI is a practical solution for product photos that get blocked by cloud content filters or can't be shared externally for privacy reasons. Here is a step-by-step guide to the required hardware, software architecture, and key considerations.

Why go local?

Cloud-based generative tools struggle on two fronts. First, content filters: requests for lingerie, swimwear, or apparel requiring models are frequently rejected. Second, privacy: uploading photos of unreleased products to third-party servers goes against the internal policies of most companies. With a local setup, images are generated on the company's own hardware and never leave the premises.

AspectCloud toolLocal setup
DataImages are uploaded to the provider's serversImages stay on the company's computers
Content filtersBound by the provider's policiesCompany sets its own rules
CostPay-per-use or subscriptionHardware investment and electricity
SetupReady to use immediatelyRequires hardware, model, and workflow setup
ConsistencyModel updates may change the outputModel and template versions can be locked

Required hardware

  • NVIDIA graphics card; at least 12 GB VRAM for smooth operation (smaller models and lower resolutions are possible with 8 GB).
  • 32 GB of system memory and at least 100 GB of free SSD space for models.
  • Good cooling if you plan on batch processing: the queue can run for hours.

Software architecture

  • ComfyUI under the hood: node-based, repeatable workflows.
  • Brand-specific style templates: lighting, angles, and background aesthetics are baked into the template.
  • Steps that preserve product shape: using a product mask to generate only the background and environment.
  • A clean desktop interface for the team: template selection, batch variation queue, and export settings.
Template selection screen of the desktop application

To enable production without requiring technical knowledge, the interface must hide the complexity of ComfyUI. Here is the application I built using this approach: Local AI Product Image Generation.

NOTE

Check the license of the model you plan to use: some open-weight image models do not allow commercial use or require a separate license.

What defines the output: templates

Choosing the right model is important, but for a consistent catalog, the real deciding factor is accurately translating the brand's visual language into templates. Start with a few approved reference images, tune the templates to match them, and test with a small batch for each new product category. If the product's color or texture gets distorted, exclude the product area from generation.

If you plan to combine image generation with text models (e.g., automated product descriptions), my note on getting JSON output from local LLMs will come in handy.

  • Shoot the product photo under consistent lighting against a plain background.
  • Extract the product mask, then generate the environment and background based on the template.
  • Queue the variations and let the team pick the best ones.
  • Export the selected images after a color check.

Quality control checklist

When generating product images with local AI, a human should always do the final check. Checking these points for each catalog image is sufficient:

  • Are the product's color and texture identical to the actual item? Compare on a color-calibrated monitor if necessary.
  • Are logos, labels, and text intact without distortion?
  • Are the shadow and lighting directions consistent with the environment?
  • If a model is used, are there any artifacts on the hands, fingers, or face?
  • Do the resolution and cropping meet the marketplace and website requirements?
WARNING

Do not use a generated image if it misrepresents the product. A significant portion of customer complaints and returns happen because the actual product looks different from the photo.

Division of labor within the team

The person who sets up the system manages the templates and models, while the team handling daily production simply uses the interface. If every change made to the templates is saved with a version number, you can easily reproduce the exact same look from a past catalog.

Conclusion

With the right hardware and well-crafted templates, generating product images with local AI lets you build catalogs without running into cloud filters or privacy concerns. For setup and team training, check out my local AI setup service.

More notes

All notes →