Unverified50% confidenceBenchmarkExact time
报告测出DiffusionGemma平均每步生成token数达到19.74个，而自回归方法每步只能生成1个
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/27/2026
First Seen
Valid until: 11/25/2026
Sources
Related Entities
Related Claims
Unverified在同一张英伟达H100显卡上，DiffusionGemma每秒可生成约1500个token，而上一代自回归模型仅能生成约300个，快了近5倍66% similarUnverifiedDuring diffusion model inference, sampler algorithms typically require only 20 to 30 iterations to obtain high-quality image generation results, far fewer than the 1,000 steps used during training.62% similarVerifiedIf the accuracy per step is 95%, after 20 steps the overall accuracy will drop to approximately 36%.62% similarUnverifiedThe ratio of input to output tokens surged from 4.52:1 to 13:1.60% similarUnverifiedMeta's LLaMA paper disclosed that approximately 67% of its training data came from Common Crawl, of which only about 15% of the original data was retained after multiple rounds of filtering.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/807341API
curl https://kongchang.com/api/v1/knowledge/claims/807341MCP
get_claim(id=807341)