Just like many of you, I’ve been playing around with ZIT these past few days. Its inference speed is really impressive. It stays lean, moves fast, and still turns out good-looking images while following prompts really well…
Since my laptop has only 8 GB VRAM, I tested several different model versions to see which works best in that constraint. Here’s a summary of what I found — in case anyone else is hesitating over which version to pick. At the end I also attached the workflow files I used, for anyone who wants them.
Since I'm not very smart and I can’t handle complicated workflows, and I also have a bit of “clean-UI OCD” where messy interfaces easily give me a headache, I just stick to the simplest, most basic workflows and try to keep everything looking as clean as possible.
Ⅰ. Versions I Tried
On my 8 GB-VRAM 5060 laptop I successfully ran these versions:
- zimageturbobf16 (11.4 GB) + qwen3_4b (7.49 GB)— hereafter “BF16”
- z-image-turbo-fp8-e4m3fn (5.73 GB) + qwen34b (7.49 GB) — hereafter “FP8”
- z-image-turbo-Q80 (6.72 GB) + Qwen3-4B-UD-Q8K_XL (4.7 GB) — hereafter “GGUF Q8”
- z-image-turbo-fp8-aio (9.63 GB) — hereafter “AIO-FP8”
- z-image-turbo-bf16-aio (19.1 GB) — hereafter “AIO-BF16”
One of the biggest advantages of the AIO versions is you only need to download a single model file. (Civitai link: https://civitai.com/models/2173571/z-image-turbo-aio)
Ⅱ. Environment & Basic Settings
- Hardware: RTX 5060 Laptop 8G VRAM, 32 GB RAM
- Software: Windows 11 Pro 25H2, NVIDIA-SMI 577.09, CUDA v12.9, Torch v2.8.0, ComfyUI v0.3.76
* T2I settings:
Size: 1024×1024
Sampler: eulercfgpp
Scheduler: simple
Steps: 9
Seed: 123456
Ⅲ. Memory Usage
As you can see, on the laptop all versions offload part of the model into system RAM, which becomes a major factor slowing down generation. VRAM usage was similar across the board: with only 8 GB VRAM, the max you can use is around 6.x GB.
If you’re planning to buy a new laptop for local T2I deployment — try to get one with more than 8 GB VRAM if your budget allows.
Ⅳ. Inference Speed
Except for the GGUF version, all the versions ran at about the same speed. From what I saw, the GGUF version is not a good match for 50-series GPUs.
Surprisingly, the AIO-FP16 version is slightly faster than AIO-FP8. At first I thought it was just measurement noise, but after repeated tests, AIO-FP16 consistently edged out AIO-FP8 by a bit.
Ⅴ. Image Quality Comparison
Based on my testing over the past few days, I feel ZIT’s image quality still has room for improvement. Sometimes the output looks “meh” — as if a small, low-quality image was simply stretched or like a LoRA trained on a rough dataset. In short: it doesn’t always look “polished,” and sometimes the focus feels a bit off. My impression is the training set for ZIT probably contains a lot of low-quality or over-processed photos from social media. Not sure if others feel the same.
On the flip side — that kind of “slightly rough” quality can also make results feel more “real,” less “AI-generated.”
I also found that longer, more detailed prompts generally produce better results than very short ones. And the choice of sampler makes a big difference. eulercfgpp turned out to be a good pick.
Here are the comparison results for three sets of short prompts:
From my personal taste — for realistic human images I’d rank them from best to worst:
BF16 ≥ AIO-BF16 > AIO-FP8 > FP8 > GGUF Q8.
BF16 and AIO-BF16 are nearly indistinguishable — basically a tie for first. GGUF Q8 was the worst: in one test, the hair region looked noticeably noisy and flawed.
For anime-style characters, again BF16 and AIO-BF16 were best. The others showed mixed results — each with their strengths and weaknesses.
Ⅵ. Conclusion
This is just my personal impression, based on a small number of samples — take it with a grain of salt. It’s totally possible that others will get very different results. That’s normal.
- If your VRAM is ≥ 8 GB or your RAM is ≥ 32 GB, then among the five models above BF16 and AIO-BF16 should be your top choices.
- AIO-FP8 and FP8 can be secondary options.
- For a 5060 laptop like mine, GGUF Q8 hardly offers any advantage.
In terms of inference speed, all five versions are noticeably slower than the FLUX 1D Nunchaku version I used before. That only makes me more eager to try a Nunchaku-based ZIT — maybe that will bring back that blazing-fast feel.
Hope this helps anyone who’s trying to pick the right ZIT version for their setup!
Have fun!
