# Generative AI Image & Video Lab

> **Status:** Part of the Scorpia studio  
> **Repository:** Private  
> **Tech Stack:** Stable Diffusion (Forge / A1111 WebUI), Realistic Vision, LoRAs, ComfyUI / ComfyDeploy, Qwen-Image-Edit, FLUX.2-klein, ControlNet, RunPod, Google Colab

---

## 1. Overview
The image engines behind the Scorpia Batch Image Studio. One studio interface can route a job to a free local Stable Diffusion install, a self-hosted Qwen-Image-Edit / FLUX GPU server, ComfyDeploy workflows on RunPod, or a Gemini cloud endpoint.

## 2. The Problem
- Paid image APIs get expensive when a catalog needs hundreds of images.
- Results drift in style from one shot to the next, which breaks a brand look.
- Restyling video (new outfit, new character) normally means reshooting.

## 3. What I Built
- **Stable Diffusion provider:** text-to-image and img2img through the Forge WebUI API, run locally at zero cost.
- **Realistic Vision + LoRAs:** a photoreal checkpoint combined with custom LoRA adapters so every product or portrait shot keeps the same look.
- **Qwen-Image-Edit server:** a Python CUDA server for instruction-based edits, with LoRAs and FLUX.2-klein.
- **ComfyDeploy on RunPod:** generate and edit workflows for jobs that need bigger GPUs.
- **AI video edit (Colab):** cuts a clip into frames, tracks the body pose on every frame and writes a skeleton video that ControlNet uses as a motion guide.

## 4. Results
- Catalog-scale batches with one visual instruction applied across a whole shoot.
- A free local path for experiments, with GPU cloud only when needed.
