NitroFusion 2 generates images in just one step
Researchers just open-sourced NitroFusion 2, slashing image diffusion down from 8 steps to just 1 or 2 steps. By introducing adversarial distillation directly into the multimodal text stream, it preserves fine details and legible typography even in a single pass. The weights are available on Hugging Face and run locally via ComfyUI.
NitroFusion 2 is a new open-source distillation project that condenses multimodal diffusion down to just one or two steps, turning sub-second image generation into an everyday reality.
Built on top of Alibaba's Z-Image architecture, the model takes a system that typically needs around 8 steps even in its Turbo variant and slashes that compute requirement dramatically, without melting text or fine details in the process.
What it can do
Traditional few-step diffusion setups generally stay within the 4 to 8 step window. Pushing them down to a single step almost always results in blurry artifacts, garbled typography, and ignored prompts.
NitroFusion 2 makes one-step sampling genuinely usable. At a single step, it renders sharp text typography and retains intricate fine-grained textures alongside accurate prompt adherence.
Give it two steps, and its output quality becomes competitive across evaluated benchmarks against models using significantly more sampling steps. It turns single-pass diffusion from a raw benchmark flex into something you can actually build workflows around.
How it works under the hood
Conventional adversarial distillation monitors generated image pixels to train the discriminator. The team behind NitroFusion 2 realized this approach ignores how modern Multimodal Diffusion Transformers (MMDiTs) actually work.
In an MMDiT, the text stream interacts directly with image tokens through joint attention, meaning the text states continuously read image data while staying tied to your prompt.
NitroFusion 2 taps into this with Text-Stream Adversarial Distillation (TSAD), using intermediate text-token representations for adversarial loss. Paired with Aspect-Cued Discrimination (ACD), which cues the discriminator to focus specifically on typography, style, or specific objects, the model learns to prioritize crucial details even under extreme step budgets.
How to run it locally
The code and weights are completely open. You can run the model locally using Diffusers or drop it straight into your ComfyUI setup.
The full BF16 model on Hugging Face weighs in around 12GB. The developers have also provided smaller compressed variants, meaning you do not need an enterprise server to test it out; mainstream consumer GPUs can handle it comfortably.
Setup instructions and integration guides are available on the project's GitHub repository linked from the project site.
Why it matters
If you run ComfyUI locally or care about interactive real-time generation, NitroFusion 2 is well worth a test run today. Cutting inference down to one or two steps transforms rapid prompt iteration and responsive UI tools.
What makes this project clever is its architecture-first insight: instead of fighting loss functions purely in pixel space, it steers distillation directly through the MMDiT text stream. Single-step generation usually sacrifices text legibility first, but NitroFusion 2 shows that picking the right representation keeps both speed and crisp typography intact.
Sources (1)
- chendaryen.github.io(opens in a new tab)Originalchendaryen.github.io