Run Z-Image-Turbo (11B DiT) in <1GB VRAM via PhobosStream

#167

by martianacdemy - opened 3 days ago

Hi everyone,

Z-Image-Turbo is an incredible single-stream Diffusion Transformer, but standard 1024x1024 loading requires ~23.5 GB VRAM when combined with the 7.5GB Qwen text encoder and attention buffers.

We open-sourced PhobosStream and ComfyUI-PhobosStream to stream DiT weights block-by-block directly from NVMe SSD on consumer hardware:

🔗 GitHub: https://github.com/martianacademy/phobos-stream
🎨 ComfyUI Custom Node: https://github.com/martianacademy/ComfyUI-PhobosStream

Benchmark on Apple Silicon Metal GPU:

Pure sampling latency: 4.79s / step (8 steps = 38.33s @ 1024x1024 resolution)
Peak memory footprint: ~920 MB (down from ~23.5 GB monolithic baseline)
Enables running Z-Image-Turbo on 8GB/12GB RTX GPUs and 16GB Macs without OOM crashes.

Feedback and PRs welcome!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment