Turbo, schnell, distilled: the fast versions of a model promise the same pictures in a fraction of the time. I wanted a number, not a promise, so here are pairs I timed on the same PC: RTX 5060 Ti 16 GB, 32 GB of RAM, 1024×1024, same settings as the official ComfyUI templates.
The pairs
| Model | Steps | CFG | Time per image | Peak VRAM |
|---|---|---|---|---|
| Z-Image (base, 16-bit) | 30 | 4 | 78.0 s | 15.5 GB |
| Z-Image Turbo (16-bit) | 8 | 1 | 11.0 s | 14.6 GB |
| FLUX.1 dev (GGUF Q8_0) | 20 | 1* | 40.6 s | 15.1 GB |
| FLUX.1 schnell (GGUF Q8_0) | 4 | 1 | 9.4 s | 15.1 GB |
* FLUX.1 dev runs at CFG 1 with its own built-in guidance (3.5). Times are sampling plus VAE decode, the average of 3 runs after a warm-up.
So Z-Image Turbo was 7.1 times faster than the base model, and schnell 4.3 times faster than dev.
Where the speed comes from
Two things, and you can almost do the maths yourself.
Fewer steps. Each step costs about the same, so 8 steps instead of 30 is already almost 4 times less work.
CFG 1. With CFG above 1, ComfyUI runs the model twice per step: once with your prompt and once with the negative prompt. At CFG 1 it runs once. Z-Image base at CFG 4 does 30 × 2 = 60 passes, Turbo does 8. That is 7.5 times less, and I measured 7.1. The rest is the VAE decode and other fixed work that does not shrink.
FLUX.1 dev already runs at CFG 1, so there only the steps count: 20 against 4 is 5 times less. I measured 4.3.
What it does not save
Memory. The fast version is the same size as the full one. Z-Image Turbo peaked at 14.6 GB and the base model at 15.5 GB; schnell and dev both at 15.1 GB. If a model does not fit on your card, the Turbo version will not fit either. Look for a smaller file instead: which GGUF to pick.
The negative prompt. At CFG 1 there is no second pass, so a negative prompt does nothing. If you rely on one, the full model is the one to use.
Which one I would start with
For trying out prompts: the fast one. At 9 to 11 seconds an image you can test ten ideas in the time the full model makes one or two. When a prompt works, run it again on the full model if you want the extra detail or need a negative prompt.
Two more fast ones I timed, without a full version to compare: FLUX.2 klein 4B made an image in 2.9 s (4 steps), and Krea 2 Turbo as INT8 in 11.0 s (8 steps).
All four workflows from the table are free to download: free ComfyUI workflows. Every other run I did is on the benchmarks page.