←back to thread

Z-Image: Powerful and highly efficient image generation model with 6B parameters

(github.com)
396 points doener | 1 comments | 30 Nov 25 11:36 UTC | HN request time: 0s | source
Show context
vunderba ◴[06 Dec 25 17:36 UTC] No.46175068[source]▶
I've done some preliminary testing with Z-Image Turbo in the past week.

Thoughts

- It's fast (~3 seconds on my RTX 4090)

- Surprisingly capable of maintaining image integrity even at high resolutions (1536x1024, sometimes 2048x2048)

- The adherence is impressive for a 6B parameter model

Some tests (2 / 4 passed):

https://imgpb.com/exMoQ

Personally I find it works better as a refiner model downstream of Qwen-Image 20b which has significantly better prompt understanding but has an unnatural "smoothness" to its generated images.

replies(6): >>46175104 #>>46175331 #>>46177028 #>>46177043 #>>46177543 #>>46178707 #
tarruda ◴[06 Dec 25 22:12 UTC] No.46177028[source]▶
> It's fast (~3 seconds on my RTX 4090)

It is amazing how far behind Apple Silicon is when it comes to use non- language models.

Using the reference code from Z-image on my M1 ultra, it takes 8 seconds per step. Over a minute for the default of 9 steps.

replies(3): >>46177803 #>>46180602 #>>46183922 #
p-e-w ◴[06 Dec 25 23:57 UTC] No.46177803[source]▶
The diffusion process is usually compute-bound, while transformer inference is memory-bound.

Apple Silicon is comparable in memory bandwidth to mid-range GPUs, but it’s light years behind on compute.

replies(1): >>46178177 #
tarruda ◴[07 Dec 25 00:47 UTC] No.46178177{3}[source]▶
> but it’s light years behind on compute.

Is that the only factor though? I wonder if pytorch is lacking optimization for the MPS backend.

replies(1): >>46180929 #
1. rfoo ◴[07 Dec 25 11:20 UTC] No.46180929{4}[source]▶
This is the only factor. People sometimes perceive Apple's NPU as "fast" and "amazing" which is simply false.

It's just that NVIDIA GPU sucks (relatively) at *single-user* LLM inference and it makes people feel like Apple not so bad.