A fully local, Apple-Silicon-optimized stack for generating photorealistic AI persona media
Photorealistic AI persona media, generated entirely on-device.
A fully local pipeline for generating ultra-realistic AI persona photos and video, wrapped in both a desktop control center and a real-time web dashboard that orchestrate a dozen chained ComfyUI workflows and LoRA training. Everything runs on-device, free, on Apple Silicon.
A look at the interface.
Representative UI — abstracted, never a client's real data.- Python
- ComfyUI
- PyQt6
- Flask
- PyTorch/MPS
Producing consistent, photorealistic AI persona content usually means paid cloud services and fragmented tooling. This stack runs entirely on-device with FP8 quantization to fit in 18GB of RAM, chaining a dozen custom nodes into reproducible photo and video workflows with one-click control.
How it
works
A dozen custom nodes chain into reproducible photo and video workflows, with FP8 quantization keeping 13GB of models stable in 18GB of RAM. A native desktop control center and a real-time web dashboard drive the same backend; tiered video graphs add temporal smoothing, audio lip-sync, and 60fps interpolation. A batch generator auto-ranks 100+ photos against a 156-template library.
What I built
(03) Inside the buildDesktop control center
A PyQt6 app manages photo enhancement, LoRA training, and generation, launching ComfyUI subprocesses and monitoring system resources.
Real-time web dashboard
A Flask + WebSocket app streams live job status and resource metrics while driving the ComfyUI executor.
Tiered video workflows
Base/Pro/Ultra graphs add temporal smoothing, audio lip-sync, 60fps interpolation, and identity-consistent 4K headshots.
The
receipts
- 12 chained ComfyUI nodes running fully local and free on Apple Silicon
- Batch generator auto-ranks 100+ photos against a 156-template library
- Two front-ends, native desktop and a web dashboard, over one backend
- FP8 quantization keeps 13GB of models stable in 18GB of RAM