%20(2).avif)
%20transparent.avif)
NitroFusion 2: AI Images in One Step on Your Own PC
NitroFusion 2 made a finished 1024×1024 picture on my laptop in 1.8 seconds, in one step, inside Promptus. The authors published pictures but no speeds, so I measured them myself: install time, seconds per picture and graphics-card memory, next to Z-Image Turbo, the model it learned from.
Every number below was measured on my own laptop (NVIDIA RTX 5090 Laptop GPU, 24 GB, ComfyUI 0.43) with seed 42 and the same five prompts. Your times will differ with your GPU.
Want to skip the setup? I saved the whole thing as a public CosyTemplate: NitroFusion 2 – One Step Image Generation.
What is NitroFusion 2?
NitroFusion 2 is an open text-to-image model that makes pictures in one or two steps instead of the usual four to eight. It comes from Dar-Yen Chen and Yi-Zhe Song, and the paper went up on arXiv on 3 November 2026.
- Size and licence: about 6 billion parameters, released under Apache-2.0, which allows commercial use (model card).
- Built on: Z-Image Turbo, so it uses the same text encoder and VAE files.
- What's new: the training judges the model's text tokens, not just the image, which the authors say gives sharper rendered text and detail in a single step.
How can one step work? Normal image models walk down a staircase: noise, a bit less noise, and so on until there's a picture. Distillation trains a fast student model to copy a slow teacher and jump straight to the bottom. NitroFusion 2 learned that jump from Z-Image Turbo, which uses 8 steps in my Promptus template.
How to install NitroFusion 2 in Promptus
- Update Promptus to the latest version first. Sharing the workflow as a CosyTemplate only worked for me after updating.
- Open the NitroFusion 2 CosyTemplate. It links to the official files:
nf2-bf16.safetensors(12.3 GB), theqwen_3_4btext encoder and theaeVAE. - If you already use Z-Image Turbo, you already have the text encoder and VAE, so only the 12.3 GB model is new. On my connection it downloaded in about 4 minutes.
- Make sure the ComfyUI server is running in Promptus's Server page, then press Run.
NitroFusion 2 speed and VRAM: my results
One step was the fastest at 1.8 seconds per picture. Z-Image Turbo took 8.4 seconds at 8 steps, so NitroFusion 2 was about 4.6 times faster.
| Version | Steps | First picture (models unloaded) | Each picture after that |
|---|---|---|---|
| NitroFusion 2 | 1 | 11.0 s | 1.8 s |
| NitroFusion 2, no LoRA | 2 | – | 2.3 s |
| NitroFusion 2, official 2-step (LoRA 0.5) | 2 | 14.5 s | 10.3 s |
| Z-Image Turbo (Promptus template) | 8 | 22.9 s | 8.4 s |
1024×1024, seed 42, same five prompts. GPU memory in use peaked at about 20.9 GB of 24 GB for both models, with the model and text encoder loaded. The very first run after downloading took 34.8 s because the 12.3 GB file was being read from disk for the first time.
Can NitroFusion 2 write text in images?
Yes, in my test. With the prompt “a ceramic coffee mug on a wooden desk, a sign on the mug that reads PROMPTUS”, NitroFusion 2 spelled PROMPTUS correctly in one step. Z-Image Turbo, with the same prompt and seed at 8 steps, wrote “PROMIPTUS”. On the ramen-street prompt, NitroFusion 2 also kept the rain and the night, which Z-Image Turbo didn't.
NitroFusion 2: 1 step vs 2 steps
The official two-step workflow adds a de-distillation LoRA (a small add-on file) at strength 0.5 on the first step and turns it off for the second; the authors recommend this to reduce artifacts. In ComfyUI on my laptop it was much slower: 10.3 seconds per picture, slower even than Z-Image Turbo. Two steps without the LoRA took only 2.3 seconds, so the LoRA was costing about 8 seconds per picture. The two-step pictures had stronger contrast and a closer, more detailed fisherman portrait. For ideas and quick tests, one step is the sweet spot.
NitroFusion 2 GGUF for low-VRAM GPUs
NitroFusion 2 also comes as GGUF files, squeezed versions that are smaller and a little less exact, from Q2_K_4XS at 2.0 GB up to Q8_0 at 7.2 GB. They need the ComfyUI-GGUF custom node. I haven't measured them yet, so I'm not giving numbers for them here.
Mistakes I hit (and the fixes)
- ComfyUI looked on but wasn't. The server had stopped while the button still said it was running. Stopping and starting it again fixed it.
- Judging speed on the first run. The first picture includes loading the model. Time the second one.
- The two-step LoRA is slow in ComfyUI. Try two steps without it before deciding.
Frequently asked questions
Can NitroFusion 2 be used commercially?
The NitroFusion 2 weights are released under Apache-2.0, which allows commercial use. Check the model card for the base model's terms too.
How much VRAM does NitroFusion 2 need?
On my 24 GB laptop GPU, memory in use peaked at about 20.9 GB with the full bf16 model and text encoder loaded. I only tested one machine, so treat that as a guide, not a minimum.
How fast is NitroFusion 2?
1.8 seconds per 1024×1024 picture after the first load, on an RTX 5090 Laptop GPU.
