[Template Bug]: z-image template generates noisy black image #433
Description
Confirmation
- I have confirmed this is a workflow template mistake
Template Name
image_z_image_turbo
Problem Description
I’m running ComfyUI v0.5 on an M4 16GB MacBook Air, using GGUF models of z-image-turbo and Qwen3-4 for fast inference.
However, I’ve noticed that z-image-turbo-Q4_K_M/S produces noisy outputs—though it works perfectly on my NVIDIA GPU-equipped PC. Only the Q3 quantized version runs reliably on macOS.
ComfyUI Mode
Legacy Mode
Error Message / Logs
no error
loaded completely; 95367431640625005117571072.00 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 327.71 seconds
got prompt
gguf qtypes: F32 (245), F16 (24), Q4_K (150), Q5_K (30), BF16 (4)
model weight dtype torch.bfloat16, manual cast: None
model_type FLOW
unet missing: ['norm_final.weight']
Requested to load Lumina2
loaded completely; 95367431640625005117571072.00 MB usable, 4526.45 MB loaded, full load: True
0%| | 0/9 [00:00<?, ?it/s]Interrupting prompt ca3d3657-983d-4ea8-9432-5cee46a4f662
0%| | 0/9 [00:09<?, ?it/s]
Processing interrupted
Prompt executed in 14.07 seconds
got prompt
Using split attention in VAE
Using split attention in VAE
VAE load device: mps, offload device: cpu, dtype: torch.bfloat16
100%|██████████| 9/9 [03:39<00:00, 24.37s/it]
Requested to load AutoencodingEngine
Unloaded partially: 2223.71 MB freed, 2302.75 MB remains loaded, 23.73 MB buffer reserved, lowvram patches: 0
Unloaded partially: 1880.86 MB freed, 421.89 MB remains loaded, 75.00 MB buffer reserved, lowvram patches: 0
0 models unloaded.
loaded completely; 95367431640625005117571072.00 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 225.67 seconds