Consistent Distribution Matching: data-free diffusion distillation beats DMD2 with 20GB VRAM
vector-institute · hf · 2026-10-19
Researchers at UBC and the Vector Institute propose Consistent Distribution Matching (CDMD), a simulation-free, data-free distillation method for diffusion and flow models.
Key contributions:
- Unifies generation and score estimation by aligning the marginal probability paths of teacher pt and student qt, requiring no curated distillation data and no computationally intensive simulations
- On ImageNet 256×256 one-step FID-50, it outperforms DMD2 and FreeFlow using 20GB of training memory on a single A100 80GB, versus 33/37 GB for baselines
- Achieves strong one-step and few-step generation results
Practically relevant for distillation under tight compute and memory budgets.
More from Multimodal
- Full AI virtual influencer character sheet: the complete template for building a creator persona — CurieuxExplorer · 2026-10-19
- Midjourney v6.1 still delivers: sref + raw parameter recipe for bold comic-style portraits — michaelrabone · 2026-10-19
- 12-task edit benchmark: smallest FLUX.3 Klein lands just 1 point behind Qwen Image 2.1 — Diligent-Rub-2113 · 2026-10-19
- How Mission Pandora keeps viewers watching a 35-minute AI film: log-entry storytelling — lmoroney · 2026-10-19
- Topview open-sources 3D Builder for blocking shots before AI video generation — Aiden_Tech_Ai · 2026-10-19
- Midjourney dual sref combo yields cohesive Lord of the Rings pen-sketch style — sergeantsref · 2026-10-19