pagedMark: Invisible SynthID watermark removal optimized for Apple Silicon
d0ofz · reddit · 2026-08-18
pagedMark is a tool optimized for Apple Silicon to remove invisible SynthID-class watermarks from AI images (ChatGPT, DALL·E, Gemini, etc.) and strip visible vendor labels. The author shares engineering insights gained while optimizing the pipeline for a 16GB M5 chip:
- LoRA Distillation Artifacts: A 4-step distilled LoRA invents textures during low-strength edits, and more steps worsen the artifact. The undistorted base model offers better fidelity at roughly 3x the wall time.
- Memory Management: On MPS, exceeding memory triggers silent swapping rather than errors. VAE Tiling is crucial to reduce peak memory (18.74 GiB -> 10.92 GiB), though it introduces faint boundary textures, requiring per-frame decision logic.
- 8GB Viability: Sequential CPU Offload streams weights to keep peak device memory at 0.28 GiB (vs 7.70 GiB resident), enabling 8GB Macs to run the model stack with acceptable latency trade-offs.
More from Infra
- Qwen 3.8 27b on a single GPU sparks debate on data center obsolescence — francoisfleuret · 2026-08-18
- YC Startup Codag Launches to Compress Agent Context 3x and Monitor Tool Usage — ycombinator · 2026-08-18
- Kubeflow graduates from CNCF, cementing a standard for cloud-native AI ops — unixterminal · 2026-08-18
- DuckDB 2.0 with Async I/O shaping up to be the fastest data lake engine — josh_wills · 2026-08-18
- Vector databases outperform note files for agent memory at scale — rseroter · 2026-08-18
- Grid Bottlenecks Stall AI: Interconnection Queues Surge to 45 Months — PeterDiamandis · 2026-08-18