NitroFusion2 advances adversarial distillation for MMDiTs with text-stream approach

Researchers have introduced NitroFusion2, a novel approach to adversarial distillation for multimodal diffusion transformers (MMDiTs). This method leverages intermediate text-token states within the transformer's text stream as a new adversarial representation, moving beyond traditional image-based discrimination. NitroFusion2 incorporates Aspect-Cued Discrimination (ACD) to steer the adversarial training process by conditioning the discriminator on specific aspects like typography, objects, or style, enabling few-step distillation down to 1-2 steps with comparable quality to conventional 4-8 step methods. AI

IMPACT This research could enable faster and more efficient generation of high-quality images from multimodal diffusion models.

RANK_REASON The cluster contains a research paper detailing a new method for AI model distillation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Aspect-Cued Discrimination
• Hugging Face
• Multimodal Diffusion Transformers
• NitroFusion
• NitroFusion2
• Text-Stream Adversarial Distillation

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →