Unverified50% confidenceBenchmarkExact time
DiffusionGemma每次生成最多允许跑48步，但通过自适应停止机制平均只需约12轮完成一次完整生成
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/27/2026
First Seen
Valid until: 11/25/2026
Sources
Related Entities
Related Claims
UnverifiedThe Speed LoRA in Bernini can compress the sampling process that originally requires 50 steps to within 20 steps through consistency distillation.65% similarUnverifiedThe latency overhead of the second-stage reranking is approximately 50–200 milliseconds, depending on the number of candidates and the model size.61% similarUnverified基准测试取三次30秒试验的中位数，交替执行顺序，丢弃预热阶段，两侧均使用单线程裸步进循环59% similarUnverifiedIn Stable Diffusion tests, Spark achieved approximately 2.9 iterations/second while Halo achieved only 1.3; for WAN 2.2 video generation of a five-second clip, Spark took approximately five and a half minutes while Halo required 75 minutes.59% similarUnverifiedTimestep distillation can compress the denoising process, which originally requires dozens of steps, down to 4–8 steps to significantly reduce inference time.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/808090API
curl https://kongchang.com/api/v1/knowledge/claims/808090MCP
get_claim(id=808090)