From Iterative to Instantaneous: Diffusion Post-Training for Real-Time High-Quality Generation
Abstract
Diffusion models are the state-of-the-art paradigm for high-quality content generation, but their slow iterative sampling incurs substantial computational costs. This thesis focuses on diffusion post-training—techniques for accelerating pre-trained diffusion models to enable real-time, high-quality, and controllable generation, breaking the inherent trade-offs between speed and quality. This work presents four core contributions. First, we propose YOSO, a self-cooperative diffusion-GAN hybrid model for one-step generation. By smoothing adversarial divergence through a mechanism where the generator learns from its own refined outputs, YOSO achieves stable training and state-of-the-art one-step text-to-image synthesis. Second, we introduce TDM (Trajectory Distribution Matching) for few-step generation, unifying trajectory distillation and distribution matching by aligning the student’s trajectory with the teacher’s at the distribution level. This data-free approach achieves strong training efficiency and surpasses the multi-step teacher. We further develop TDM-unify for flexible, deterministic sampling across variable step counts. Third, we propose Reward-Instruct, a reward-centric approach challenging the assumption that distillation losses are indispensable for few-step generators. We show that as conditions become more specific, reward signals emerge as the primary driving force, reducing distillation to costly regularization. Reward-Instruct directly converts pre-trained…
Read the paper
Related studies
- Learning from First-Generation Qualitative Approaches in the IS Discipline: An Evolutionary View and Some Implications for Authors and Evaluators (PART 1/2)
- Commentary on the clinical and preclinical dosage limits of interstitially administered magnetic fluids for therapeutic hyperthermia based on current practice and efficacy models
- Prosthesis: New Technological Opportunities and Innovative Biomedical Devices
- The history and future of neural modeling for cochlear implants
- Ryanodine receptors: waking up from refractoriness
- Novel Transgenic Rabbit Model Sheds Light on the Puzzling Role of Matrix Metalloproteinase-12 in Atherosclerosis
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.