Engineering Challenges of Removing AI Watermarks on Metal

d0ofz · reddit · 2026-08-17

A technical breakdown of running an SDXL-based pipeline to remove AI watermarks (SynthID, C2PA) on Mac Metal. Key findings: Distilled LoRAs introduce artifacts and degrade image quality; VAE tiling is essential for low-VRAM devices; Sequential CPU offload enables 8GB devices but slows performance. The author also highlights critical Metal backend limitations: lack of float8 support, float64 issues with SAM prompts, and silent zero returns with fp16 sampling optimizations.

Original post →

More from Infra

Infra channel →