NitroFusion2 open-sourced: sharp 1-2 step images with readable text via text-stream adversarial loss
pbaylies · x · 2026-11-05
NitroFusion2, a 1–2 step text-to-image model built on Z-Image-Turbo, is released under Apache-2.0. Its key trick is moving adversarial loss onto the text stream of the MMDiT — since the text stream has already read the image, it can be steered to fix typography and object details, rendering sharp readable text in a single step. Checkpoints are on Hugging Face with ComfyUI workflows (1-step and 2-step), diffusers code and an online demo.
More from Multimodal
- Midjourney sref combo produces fuzzy towel-textured surrealist faces — sergeantsref · 2026-11-06
- Dev releases pixel art model for game dev with exact size control and sprite-based style consistency — HugoDuprez · 2026-11-06
- Fan creates stunning AI-generated Terraria short film — sadelcri · 2026-11-06
- MOSS Base image model nails an absurd multi-element prompt in one shot — cgs019283 · 2026-11-06
- AI-generated cinematic Ronaldo series goes viral with GOAT throne imagery — egeberkina · 2026-11-06
- Stunning AI-generated image of a reader inside an endless green encyclopedia universe — JosephJacks_ · 2026-11-06