[Jetson AGX Orin/Edge AI] Fast, High-Fidelity Img2Img? Seeking EFFICIENT Face Preservation Solutions!

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Orin
,
1

Hi everyone,

I’m working on a university project to build an AI-Photobooth running fully on a Jetson AGX Orin (64GB). The core task is: Take a photo (Img2Img) → Apply style transform while preserving the face. I’m new to the Jetson Universe.

🚨 Direct Question: Speed & Quality on Edge Hardware

Has anyone achieved fast and efficient Img2Img processing on the Jetson without loss of facial identity?

We need near real-time performance (5s-30s), and any proven solutions for face preservation that run quickly and don’t balloon the memory footprint.

My Debugging Roadmap (Why I’m Struggling)

Here is a summary of the methods and models I have already attempted, which have led to roadblocks:

  • SDXL-Turbo / SDXL: Failed due to the size of the model and Unified Memory (VRAM/RAM) limitations on the Jetson. Or it failed because there was no functional IP-Adapter for SDXL-turbo (at least I didn’t found one)

  • SD 1.5 Base Model: Successfully loaded after optimization (torch.float16 and using device_map="balanced" to avoid OOM errors).

  • IP-Adapter Loading (via Hub): Failed repeatedly due to persistent Hugging Face Naming/Cache Conflicts, as the official pipe.load_ip_adapter() could not locate the weight file.

  • IP-Adapter Loading (Local Bypass): Failed even after manually downloading the ip-adapter-plus_sd15.bin file to a local folder .

If you have experience with optimized pipelines or alternative methods, please share your workflow!

Thank you for your help!

5

Hi,

Have you checked the “Image Generation” section in the Jetson AI Lab?

Suppose these examples meet your requirements.
Thanks.