The Nunchaku version of the ZIT model was released a couple of days ago, and today it finally finished adapting to ComfyUI. I couldn’t wait to try it out.
Short version first: it delivers the same high performance as other Nunchaku models — no disappointment here.
Based on my previous ZIT test post, I’ve now added the Nunchaku-related results and updates.
For this quick test, I’m still using my RTX 5060 laptop with 8GB of VRAM.
I. Versions Tested
According to the official documentation, 50-series GPUs should use the NVFP4 versions. I tested both NVFP4 variants released officially:
- svdq-fp4r32-z-image-turbo (3.53GB) + qwen3_4b (7.49GB)
- svdq-fp4r128-z-image-turbo (3.91GB) + qwen3_4b (7.49GB)
Official model download page: https://huggingface.co/nunchaku-tech/nunchaku-z-image-turbo
II. Environment and Basic Settings
- Hardware
- RTX 5060 Laptop, 8GB VRAM
- 32GB RAM
- Software
- Windows 11 Pro 25H2
- NVIDIA-SMI 577.09, CUDA v13, Torch v2.9.1
- ComfyUI v0.6.0
- Basic T2I Settings
- Size: 1024 × 1024
- Sampler: eulercfgpp
- Scheduler: simple
- Steps: 9
- Seed: 123456
To use the new model, I updated both ComfyUI and Nunchaku to the latest versions.
It’s worth mentioning that the error with Nunchaku’s LoRA Loader node, which appeared after ComfyUI was updated to v0.4.0 a few days ago, has now been fixed. Also, the newer versions of ComfyUI noticeably improve startup speed compared to before.
As for upgrading Nunchaku, you may run into some weird or unexpected issues. I’d strongly recommend doing a clean reinstall instead of upgrading in place. There are two parts to this:
1. Nunchaku ComfyUI custom nodes
Simply delete the Nunchaku folder under ComfyUI/custom_nodes.
2. Nunchaku core package
Uninstall it via pip.
After everything is removed, download the latest Nunchaku core from the official GitHub repo (currently v1.1.0 as of today). Make sure to choose the version that matches your setup.
The Nunchaku ComfyUI plugin can then be installed easily via “Manage Extensions” inside ComfyUI.
Official Nunchaku GitHub: https://github.com/nunchaku-tech/nunchaku
III. Memory Usage
The Nunchaku version uses roughly the same amount of memory as the other versions, with slightly lower VRAM usage. On an 8GB VRAM laptop, it runs without any issues and feels quite comfortable overall.
Ⅳ. Inference Speed
No need to say much here — Nunchaku is still way faster than other models at this point.
The r128 version is just a tiny bit slower than r32, but with only 9 steps, the total generation time is almost the same.
For reference, on the same laptop I previously tested the Nunchaku FLUX version. It ran at about 1.02 it/s, generating 1024×1024 images at 20 steps, using the same eulercfgpp + simple setup, with each image taking around 19.8 seconds.
With Nunchaku ZIT, both the r32 and r128 versions take roughly 12 seconds per image. For an entry-level laptop like mine, that’s already very impressive performance.
V. Image Quality Comparison
I’m still using the same three image comparison sets from my previous post, and I’ve now added two Nunchaku samples as well.
(The sample images below can be opened in a new window via the right-click menu to view them at full resolution.)
From a personal preference standpoint, the r128 version looks very close to the BF16 version in terms of image quality, and the r32 version is also completely acceptable. Overall, Nunchaku maintains its usual quality level here and doesn’t disappoint.
Ⅵ. Conclusion
If you’re still on the fence about whether to go with Nunchaku — especially if you’re using an entry-level GPU like I am — there’s really no reason not to choose the Nunchaku version of ZIT.
For RTX 50-series users, the r128 version is the top recommendation.
Just like with Nunchaku FLUX, Nunchaku ZIT is going to be my main go-to model on my laptop.
Now all that’s left is to wait for Nunchaku’s support for ZIT LoRAs. Hopefully, we won’t have to wait too long.
Since I’m using the official workflow without any modifications, I won’t repost it here. Once LoRA support is available, I’ll release a new, simplified workflow at that time.
Have fun!
