Unverified50% confidenceFactExact time
DiffusionGemma复用现成的Gemma系列（约26B）模型权重，通过监督微调和采样器蒸馏与强化学习两步改造完成转换
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/27/2026
First Seen
Valid until: 11/25/2026
Sources
Related Entities
Related Claims
UnverifiedDiffusionGemma是首个在大规模（26B参数）上证明扩散范式可以匹敌自回归质量的工作80% similarUnverifiedCurrent mainstream text-to-image models include Stable Diffusion, DALL-E, and Midjourney, while text-to-video models include Sora, the Runway Gen series, and the underlying model of Jimeng AI.69% similarVerifiedRepresentative products of mainstream Diffusion Models include Stable Diffusion, early versions of DALL-E 2, and Midjourney; their working principle involves progressively adding Gaussian noise to images during training and then training a neural network to learn the reverse denoising process.69% similarUnverified2024年以来，Mistral、Phi、Gemma等小型模型通过更精细的数据筛选、知识蒸馏和训练策略优化，在特定任务上逼近甚至超越了体量数倍于己的大模型69% similarUnverifiedFine-tuning involves taking an existing large model and further training it with a specially prepared proprietary dataset, thereby changing the model's weight values67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/807342API
curl https://kongchang.com/api/v1/knowledge/claims/807342MCP
get_claim(id=807342)