H3 is a video model, but render exactly ONE frame of R2V through Mamad8's single-image VAE and it becomes a serious instruction-based image editor. Credit for the technique goes to Patient_Ratio4177. This is the Apple Silicon port of it, since the original recipe is CUDA-only (int8_convrot + nvfp4 + comfy-kitchen).

What's in the repo (github.com/Bambushu/h3edit-mac):

- the ComfyUI workflow (GGUF Ref2VA + lightx2v turbo LoRA + ClipProj, sa_solver/8, fully MPS)

- a h3edit CLI: one command, two reference images in, one edit out

- a tiny custom node that unlocks 1-frame renders WITHOUT patching ComfyUI's source (survives updates)

- 4 worked demos with full prompts + 1 honest failure

~7-10 min per image at 2208x960 on an M5. All demo brands are fictional and AI-generated.