NitroFusion2 Released: 1-2 Step Text-to-Image via Text-Stream Adversarial Distillation

The SketchX team at the University of Surrey has released NitroFusion2, a 1-2 step text-to-image model built on Alibaba's Z-Image-Turbo, with the project page, code, paper, and model weights all available now.

Confirmed

The core technique is Text-Stream Adversarial Distillation (TSA): it shifts the discriminator in adversarial distillation from image features to the text stream of the multimodal diffusion Transformer (MMDiT). Thanks to joint attention, the text stream has already "read" the image and is bound to the text conditioning, making it a richer adversarial representation
This design lets the model render crisp text (e.g., book cover lettering) and fine detail even in a single step, with 2-step settings further improving quality
The authors' comparison images show its text rendering and portrait detail generally outperforming baselines such as TwinFlow and Z-Image in both 1-step and 2-step settings

Why it matters

Adversarial losses typically operate on image features; moving them to the text stream via TSA opens a new distillation path, offering an alternative to conventional discriminator design for compressing diffusion models into few-step ones
With image generation in 1-2 steps and clear text rendering, the model suits latency-sensitive real-world applications, and it's fully open source (code, paper, weights), making it easy for the community to reproduce and build on

2026-11-05 ~ 2026-11-06 · 5 related posts

Primary sources

NitroFusion2 ships 1–2 step text-to-image via text-stream adversarial distillation on MMDiT — bdsqlsz · 2026-11-06 00:02 (UTC+08:00)
NitroFusion2 distills diffusion transformers to 1-2 steps via text-stream adversarial loss — SketchX · 2026-11-05 19:17 (UTC+08:00)
NitroFusion2 debuts: a 1-2 step text-to-image model built on Z-Image-Turbo — bdsqlsz · 2026-11-06 00:02 (UTC+08:00)

[source] NitroFusion2 distills diffusion transformers to 1-2 steps via text-stream adversarial loss — SketchX · 2026-11-05
NitroFusion2 open-sourced: sharp 1-2 step images with readable text via text-stream adversarial loss — pbaylies · 2026-11-05
[source] NitroFusion2 debuts: a 1-2 step text-to-image model built on Z-Image-Turbo — bdsqlsz · 2026-11-06
[source] NitroFusion2 ships 1–2 step text-to-image via text-stream adversarial distillation on MMDiT — bdsqlsz · 2026-11-06
NitroFusion2: 1-2 Step T2I Model Puts Adversarial Loss on MMDiT Text Stream — Crazy-Repeat-2006 · 2026-11-06