Dev Builds ComfyUI Frontend for Qwen-Image 2.1 on Mac, Documents MPS Fixes
Spiritual-Savings377 · reddit · 2026-09-30
A developer open-sourced Image Studio, an AGPL local web app running Qwen-Image 2.1 via ComfyUI headless on Apple Silicon, replacing manual graphs with a prompt box and a small canvas of four node types.
Useful MPS tuning details for Mac users:
- --gpu-only keeps text encoders on GPU, 35% faster edits
- --cpu-vae: VAE encode/decode on MPS corrupts images (mean error 54/255 vs 1.8 on CPU)
- Official int8 text encoders rely on aten::intmm, missing in MPS; a one-time bf16 conversion lifts the 9B prompt enhancer from 0.4 to 11 tokens/s
Requires 64GB memory and Docker; weights stay under the non-commercial Qwen Research License.
More from Infra
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- US BIS Closes Remote-Access Loophole for Advanced AI Chips in New Rule — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16