类 Sora 开源架构模型训练实践-卞正达.pdf

《类 Sora 开源架构模型训练实践-卞正达.pdf》由会员分享，可在线阅读，更多相关《类 Sora 开源架构模型训练实践-卞正达.pdf（35页珍藏版）》请在三个皮匠报告上搜索。

1、OPEN-SORA潞晨科技CTO 卞正达Democratizing Efficient Video Production for AllScan me for the code!Generated By Open-SoraContentsIntroduction of OpenAIs SoraWhats Open-Sora?Technical Insights into Open-SoraPerformance Future PlanScan me for the code!Introduction of OpenAIs SoraScan me for the code!Introductio

2、n of OpenAIs SoraScan me for the code!Sora is a generative text-to-video AI model developed by OpenAI,the makers of ChatGPT and DALLE 3.“It can create realistic and imaginative scenes.”source:https:/ and PerformanceScan me for the code!source:https:/ beat Pika,Runway and Stable Video in a single nig

3、htApplications and Use CasesScan me for the code!Gaming and Virtual RealityArt and Creative ExplorationMedia ProductionSimulations for Drug DiscoveryAdvertising and MarketingEducation and TrainingZak Kukoff,an early investor in San Francisco stated that producing a movie for under$50M is now feasibl

4、eGenerated by OpenAIs SoraWhats Open-Sora?Scan me for the code!Open-Sora:the First Open Source Sora-like Video Generation ModelScan me for the code!Bringing OpenAIs Sora model to the community with low-cost,fully open-source replication:Model ArchitectureTrained Model CheckpointsTraining Process Det

5、ailsData PreprocessingVideo Demonstration and TutorialTechnical Insights into Open-SoraScan me for the code!Open-Sora:Model ArchitectureScan me for the code!Model architecture designTraining reproduction schemeData preprocessingEfficient training strategies from Colossal-AI1234Open-Sora:Model Archit

6、ectureScan me for the code!Fig.1 STDiT Model Structure SchematicUtilizes DiT ArchitectureBuilding upon the popular DiT framework and use powerful text-to-image model PixArt-as a strong initialization.Fig.2 Training Speedup with STDiTReduce Training and Inference CostSTDiT surpasses DiT in training e