20230912_方正证券_传媒行业3D研究报告:AI的下一个涌现_45页.pdf
分析师 杨晓峰 登记编号:S1220522040001杨昊 登记编号:S12205230700043D:AI的下一个涌现传媒团队 行业深度报告证 券 研 究 报 告|传媒行 业|2023 年9 月12 日 2D 3D 2D 1 GAN 2D 2 2D ImageNet 22000 1500 ImageNet ILSVRC AlexNet VGG GoogleNet 3 LAION Stable Diffusion LAION-5B 3D OpenUSD 3D OpenUSD 3D1:7 11 Objaverse-XL 1020 3D 3D 3D NeRF Objaverse-XL 3D 2D 2020-2021 2 USD 3D 3D 3D 2XVBYvMqNpMmNpRtQnQsNnMbRaObRoMqQoMnOlOmMxOjMpOrN9PnMsOvPnPmRMYpPtM 3D-+NeRF1 4 3D 3D 3D 3D 3D 2 3D 3D 3D DreamFusion 3D AI DreamFusion 3D+3D NeRF 3D3 OpenAI Meta Apple AI+3D 3D 3D 3D 3D1 2D 2D 2D 20-50 2 3D 3D AI 30000 3D-Zero123 3090 3D 3-4 3 3D 5 2 3D 2.6 3D3 3D VR/AR MR 3D3D 3D 3D 2022 4 2D 3D OpenUSD 3D-+NeRF 3D5一、文生2D复盘:“千万级数据和亿级数据”是关键1.1 2D 7 prompt LeNet AE CV LeNet Yann Lecun 2018 CNN LeNet MNIST 21 CIFAR100 AE Datawhale GPT-3 MNIST GPT-3 LeNet 8 AI Midjourney1.2 2D 2022-2022 2D 2009-2020 2020-2021 2022-2022 2022-2023 2023 2009 ImageNet 2014 12 GAN 2015 11 GAN(DCGAN)2015-2020 GAN 2021 1 DALL-E CLIP 2021 6 GAN 20212 4 DALLE 2 prompt engineering NeurIPS 2022 CVPR 2022 NeRF NeurIPS 2022 best pape Imagen LAION-5B edm 2022 12 Stable Diffusion 2.0 2023 Midjourney V4 V5 2.0/.2D 2D 9 arXiv paperwithcode 1.2.1 GAN 42%31%6%6%5%3%3%3%3%降噪 图像生成 文本到图像生成 语义分割 超分辨率图像分类 物体检测 语言建模 图像去噪 diffusion models 2D GAN NLP SOTA GAN 2014 Generative Adversarial Networks 2016 Conditional Image Generation with PixelCNN Decoders 2020 Denoising Diffusion Probabilistic Models 1.2.2 ImageNet 2D 10 paperswithcode ImageNet 09 1500 ImageNet 2007 2009 320 2D ImageNet 22000 1500 ImageNet Mechanical Turk Mechanical Turk ImageNet 167 4 9 3 2007-2010 ImageNet ImagNet 1998 CMU/VASC Faces 337 750,000 1998 FERET Faces 1199 14,126 1998 MNIST digits 70,000 1999 CuRRET Textures 5,000+2001 Middlebury Sterco 2003 CalTech 101 9,146,256 2004 KTH human action 2,391 2006 ESP 10002006 MSRC 30 12 2007 PASCAL 20 9963 2007 Lotus HILL 500,000 2007 CalTech 256 30,607,257 2008 LabelMe 2008 TinyImages 79,300,000 32*32 2009 ImageNet 300 1500 11460 2127 6256 23991 1612 4385 2357 15.3%13.5%6.7%3.6%3.0%2.3%0.0%2.0%4.0%6.0%8.0%10.0%12.0%14.0%16.0%18.0%0500010000150002000025000300002012 2013 2014 2015 2016 2017 11 CSDN ImageNet ILSVRC,2010 2017 Kaggle 2016 172 7 0.28 0.03 ILSVRC AlexNet 12 GoogleNet 14 ResNet 15 SENet 17 ImageNet AlexNet VGG 2014 GoogleNet 6 15%2%ILSVRC TOP 1 352029891231571720.280.030.230.6600.350.7040801201602002010 2011 2012 2013 2014 2015 2016 ILSVRC 1.2.2 ImageNet AlexNetZFNetGoogleNetResNetResNeXtSENet12 paperswithcode OpenDataLab LAION 2021 LAION-400M,22 10 LAION-5B 14 LAION Common Crawl LAION Imagen Stable Diffusion 2D MS-COCO 330,000 CC3M 3,000,000 Visual Genome 5,400,000 WIT 5,500,000 CC12M 12,000,000 RedCaps 12,000,000 LAION-5B 230,000,000 CLIP WIT 400,000,000 ALIGN 180,000,000 BASIC 660,000,000 Stable Diffusion StabilityAl LAION-5B DALL-E 2 OpenAl CLIP DALL-E(650M)Midjourney Midjourney Imagen Google 460M-Laion 400M&1.2.3 LAION 二、3D研究框架:已破千万级数据集,OpenUSD加速数据集扩张14 2 3D 3D 3D 3D 3DOpenUSD3D 3D OpenUSD 3D 3D 3D 3D OpenUSD 3D 3D 3D 3D Point Cloud Polygon Mesh Voxel Multi-view ImagesOccupancy FunctionSDF Signed Distance Function INRs Implicit Neural Representations x y z-SDF(R,G,B)3D 3D MLP 3D 15 SIGAI CSDN,2.1.1 3D INRs 3D NeRF 3D 3D INRs NeRF 3D16,2.1.2 USD 3D.gltf/.gllb.obj.fbx.stl.3ds usd/.usdz 3D 3D 3D 3D Web 3D 3D 3D USD 3D.usd/.usdz Web AR/VR17 OpenUSD NVIDIA USD 3D HTML 23 8 8 SIGGRAPH NVIDIA HTML 2D OpenUSD 3D AOUSD USD USD 3D USD Pixar USD OpenUSD AOUSD Adobe Autodesk Linux NVIDIA Omniverse USD 3D 2.1.3 USD 3D,OpenUSD USD NVIDIA Omniverse OpenUSD AI Omniverse Kit RTX Omniverse USD Composer Omniverse Audio2Face Omniverse Cloud API.Adobe FireflyWonder Dynamics3D 图像捕捉平台Luma AI角色引擎公司Inworld AI虚拟形象公司ConvaiBlackshark.AI 世界数字孪生平台.18 Objaverse-XL:A Universe of 10M+3D Objects 2023 7 11 Objaverse-XL 1020 3D Objaverse1.0 bjaverse-XL 3D 1020 3D 2D 3D NeRF Objaverse-XL 3D2D 2020-2021 3D 3D56%35%8%1%GitHub Thingiverse Sketchfab Polycam and the Smithsonian Institute2.2.1 3D 2D 2020-2021 Objaverse-XL Objaverse19 Objaverse-XL:A Universe of 10M+3D Objects CSDN PixelNeRF NeRF Zero123 3D PixelNeRF PSNR(Peak Signal-to-Noise Ratio)Objaverse-XL 1000 Objaverse 800 Zero123-XL Zero123 3D Zero123-XL Zero123 PixelNeRF PSNR 2.2.2 3D 20 WYlog Sketchfab 3D 3D Blender Maya3D-3D 3D 3D Sketchfab 3D$3-$500 Sketchfab 6 UV 2D 1 2 3 4 5 3D 2.2.3 3D 21 3D 3D AI+3D 3D AI+3D 3D 3D 2.2.4 3D 3D 3D3DAI+3D3D3D 3D 3D 三、文生3D的方向-扩散模型+NeRF23 BIM 3.1 3D-3D 3D 3D 3D 3D 3D 2D 3D Prompt 3D 4 3D 3D 3D 3D 3D 3D 3D 3D 24 AIRX 3D Scanner App 3.1.1 3D 3D CT AR/VR 2015 5mm 2020 iPad Pro 3D LIDAR 3D Scanner Pro 3D iOS 12 Quicklook USDZ 3D RealityScan iOS 1 2 3 A3D 3D 3D Scanner Pro 25 GGAC NVIDIA 3.1.2 3D 3D 22 Luma NeRF iPhone AI 3D 2022 10 Connect Codec Avatars 2.0 Instant Codec Avatars CYAN.AI CNN DNN 2D 3D Unity 22 10 Luma 22 11 iOS App 22 12 3D 23 1 iOS App NeRF Reshoot23 1 NeRF 23 2 NeRF 23 3 iOS App AR 3D API Luma AI Instant Codec Avatars 2023 4 Luma Unreal Engine Alpha2023 5 Unreal Engine plug-in V22023 7Unreal Engine plug-in v0.32023 8 Flythroughs26 AI GameLook 3.1.3 3D 3D 3D 3D NeRF 3D Kaedim Kaedim3D 3D 3D PIFuHD 2D 3D NPC 1 Midjourney V5 2 PIFuHD 3D 3 3ds Max 4 UV PIFuHD 27 Tafi 3.1.4 3D 3D 3dfy.ai Tafi Masterpiece Studio2023 6 Tafi 3D 3D 3D Tafi Genesis 3D 3D 3D DCC 3D Unreal Unity Blender Maya Maxon Cinema 4D 3D Tafi 4 Tafi 28 DreamFusion:Text-to-3D using 2D Diffusion 3.2.1 DreamFusion-Imagen+NeRF DreamFusion 2D NeRF 3D AI 2D Prompt(NeRF)3D DreamFusion AIImagen NeRF 3D 3D 3D NERF MLP 64 64 NeRF 15,000 1.5 Nerf 3D“a DSLR photo of a pencock on a surfboard”DreamDiffusion 29 3D 1 2 3D DreamFusion Point-E OpenAI Magic3D ProlificDreamer 2D 3D 3D 3D 3D2022 11 Magic3D LDM NeRF-Instant NGP2022 12 Point E OpenAI DALL-E2 3D 3D 2023 5 ProlificDreamer 2D Stable-Diffusion+LoRA VSD DMTet Instant NGP3.2.2+3D 3D 30:NeRF:Representing Scenes as Neural Radiance Fields for View Synthesis Dimensions,NeRF neural implicit representation 2019 2020 ECCV NeRF:Representing Scenes as Neural Radiance Fields for View Synthesis NeRF 3D NeRF 2022 NeRF 615 2020-2022 NeRF 1074 R,t NeRF NeRF3.2.3 NeRF 3D 1628761541601002003004005006007002020 2021 2022 2023 31 3.3 3D OpenAI Meta AI+3D 3D 3D 3D 23 ChatGPT 2D AI 3D:1 2D 2D 3D 2 3D 1 3D 22 Get3D Magic3D 3D 3D 3D AUSD 3D2 DreamFusion 3 Meta Facebook 3D Meta 3D VR/AR 3D4 OpenAI 23 AI GPT DALLE 3D Shap-E Point-E 3D 5 Apple 3D-AR 3D WWDC AR/MR32 2022 11 Magic3D 3D 3D 3D Instant NGP latent Instant NGP 3D Magic3D 3D 2022 9 GET3D 3D 3D AI StyleGAN-NADA 3D 3D 3D 3D3.3.1 3D-3D 33 AIRX 2023 8 Neuralangelo 3D 3D 3D 3D SIGGRAPH AI 3D AI 3D 3D,3D 3D NeRFs AI,3D,3D Neuralangelo 3D SIGGRAPH 2023 3D 3D3.3.1 3D-3D 34 CVer AIRX 2022 10 DreamFusion 3D 2D NeRF 3D2022 9 2D 3D 3D landmarks 3D 3D 3D 3D3.3.2 3D-3D35 CG Meta2023 3 Meta Make-A-Video3D 4D 3D+4D NeRF T2V 3D AR/VR 3D 2022 10 Codec Avatars 2.0 Prior model Meta MAV3D 4D iPhone VR 3D 3D3.3.3 3D-Meta 36 OpenAI2023 5 OpenAI Shap-E Point-E+NeRF INR 3D 2022 12 OpenAI Point-E 3D+Point-E 3D Shape-E Point-E Point-E 3D 3D3.3.4 3D-OpenAI 37 XR AR AR Apple Object Capture IOS 3D3.3.5 3D-Apple-WWDC 2017 WWDC 2018 WWDC 2019 WWDC 2020 WWDC 2021 WWDC 2022 WWDC 2023 ARKit AppleAR AR USDZ QuickLook ARQuickLook ARKit AR RealityKit 3D RealityConverter RealityCompos ARKit Lidar RealityConverter AppleMap 3D iOS15 AR RealityKit2 Object Capture RoomPlan,3D Vision Pro,vision OS 3D iOS 17 Object Capture 3D-AR USDZ RealityKit 3D Reality Converter Lidar Vision Pro 3D 3D AR/MR四、文生3D成本测算:迭代次数万级以上39 Stable Diffusion 4.1 2D 20-50 2D 2D 20-50 RTX 3090 Stable Diffusion 3 2D 2D 2.1-4.1:5 2.41 2.1/s20 5.7 3.5/s50 12.2 4.1/s40 4.2 3D-3-4/3D Zero123 3D 30000 RTX 3090 24GB 3.3-4.2 Zero123 30000 3D NeRF 3D 3090 24GB 50%12GB 3D 2.35=/60seconds/60minutes 30000 3D 3.3-4.2:GPU 309025000 30000 35000 400001 6.9 8.3 9.7 11.1 1.5 4.6 5.6 6.5 7.4 2 3.5 4.2 4.9 5.6 2.5 2.8 3.3 3.9 4.4 3 2.3 2.8 3.2 3.7 3.5 2.0 2.4 2.8 3.2 4 1.7 2.1 2.4 2.8 41 sketchfab 4.3 3D-5/:A100 RTX 3090 Zero123 30000 3D 5 RTX 3090 3D Zero123 RTX 3090 1.39 2.25 30000 3D=5.15 30000/2.25/s/60min/60sec*1.39/hour 2 3090 3D 2.6 3D 3D 3D sketchfab 3D 2-15 3-40 AI+3D 3D 3 2 2 15 40 650 50:sketchfab 3D/s)25000 30000 35000 40000 2.00 4.83 5.79 6.76 7.72 2.25 4.29 5.15 6.01 6.86 2.50 3.86 4.63 5.41 6.18 2X4.00 2.41 2.90 3.38 3.86 4.50 2.15 2.57 3.00 3.43 5.00 1.93 2.32 2.70 3.09 3D VR/AR MR 3D 3D 3D 3D 2022 42 43 44方正证券研究所上 海 市 静 安 区 延 平 路 7 1 号 延 平 大 厦 2 楼深 圳市 福田 区竹 子林 紫竹七 道光 大银 行大 厦31 层广 州 市 天 河 区 兴 盛 路12 号楼隽峰院2 期3 层 方 正 证 券 北 京 市 西 城 区 展 览 路48 号 新 联 写 字 楼6 层长 沙 市 天 心 区 湘 江 中 路 二 段36 号 华 远 国 际 中 心37 层专 注 专 心 专 业45