AI Graphic Design Pipeline Development

I want to build an end-to-end AI pipeline that takes a few reference images and turns them into fresh, production-ready graphic designs. Users will land on a simple web portal, drag-and-drop photographs, illustrations, or even quick screenshots, and the system will extract the visual cues that matter most—colour schemes, overall layout and composition, and typography style. Those cues will condition a generative model so the output feels inspired by the references yet remains original.

The workflow I have in mind looks roughly like this: an ingestion layer that stores and tags the user images, a feature-extraction service (likely using a vision transformer or CLIP-based model), prompt building, and finally a design generator—whether that ends up being a fine-tuned diffusion model, GAN, or another approach you recommend. A lightweight front end should let users tweak parameters, run new iterations, and download high-resolution results.

Deliverables
• Web portal (upload, preview, parameter controls)
• Backend services for image storage, feature extraction, and prompt conditioning
• Trained generative model focused on colour palettes, composition, and typography
• API or microservice endpoints so the pipeline can be expanded later
• Setup scripts and concise documentation covering installation, retraining, and model upgrade steps

I’m open to your preferred stack—Python, TensorFlow, PyTorch, FastAPI, or similar—as long as it’s reproducible in a cloud environment. Let me know which tech you’d lean on, any pre-trained models you would start from, and a realistic timeline for an MVP.