NitroFusion2 is a new open text-to-image model that needs only one or two steps per picture. It is built on Alibaba's Z-Image-Turbo and moves the adversarial training onto the text stream of the model, which keeps text in images sharp even at a single step. The weights are Apache-2.0, but running them needs a 24 GB GPU. On LiveWeights you can try it in your browser.
NitroFusion2ChenDY (University of Surrey) · Apache-2.0A 1-2 step image generator built on Z-Image-Turbo: sharp, detailed pictures with readable text in a couple of seconds.Try it online →What's new in NitroFusion2
- 1-2 steps instead of 8 or more: an image takes a few seconds, so you can iterate on a prompt quickly.
- Readable text: signs, labels and titles come out clean, even in one-step mode.
- A de-distillation LoRA: at two steps the authors blend the base model back in on the first step to remove artifacts. We use their recipe.
- Small enough to share: GGUF versions from 2 to 7 GB run in ComfyUI on modest GPUs.
License: can I use it commercially?
NitroFusion2 is released under Apache-2.0, like the Z-Image-Turbo model it builds on, so commercial use is allowed. The images you generate on LiveWeights are yours.
How to try it online
- Open the NitroFusion2 page and sign in, then pick a small credit pack (no subscription).
- Describe the image. Put any words you want written in quotes.
- Choose the format and the steps: Balanced (2 steps) is the authors' recommendation, Instant (1 step) is the fastest, Detailed (4 steps) adds a little more detail.
- Press Generate. Reuse the seed to redo the same image with different settings.
NitroFusion2 or MOSS-Image?
Both are open and run on LiveWeights. NitroFusion2 is much faster and cheaper per image, ideal for exploring ideas. MOSS-Image takes about 40 seconds but follows very long, detailed prompts and handles Chinese text.
Run it locally
The weights are on Hugging Face (ChenDY/NitroFusion2), with a diffusers inference.py, ComfyUI workflows for 1 and 2 steps, and GGUF files from Q2 to Q8. In bf16 the transformer and text encoder need about 21 GB; with CPU offload or a GGUF it fits on 12-16 GB cards.