NitroFusion2 ships 1–2 step text-to-image via text-stream adversarial distillation on MMDiT
bdsqlsz · x · 2026-11-06
NitroFusion2 released: a 1–2 step text-to-image model built on Z-Image-Turbo, with page, code, paper, and weights all public.
The core idea is Text-Stream Adversarial Distillation (TSAD): in MMDiT architectures, text tokens pass through joint attention and thus "read" the image, becoming discriminative representations still tied to the text condition. Applying adversarial loss on intermediate text-token states lets a single frozen discriminator attend to different aspects of the generated image (typography, objects, style) simply by swapping the conditioning cue — one backbone, many critics, no extra cost.
More from Multimodal
- Full prompt for a 30-second AI drama scene of feuding neighbors — LudovicCreator · 2026-11-06
- Seedance 2.5 short film: neighbors fight over leaves, prompt shared — LudovicCreator · 2026-11-06
- OpenAI's Justin Uberti to reveal GPT-Live speech-to-speech architecture at VapiCon 2026 — juberti · 2026-11-06
- Pruna open-sources P-H3-Image: MiniMax-H3 video model turned into 8-step, 7x smaller local image model — cocktailpeanut · 2026-11-06
- Melite open-sources agentic ComfyUI: agents that generate and extend their own harness — data-cypher-000 · 2026-11-06
- HeyGen gets 50% cheaper via Monid, no subscription needed for AI avatar videos — socialwithaayan · 2026-11-06