朱昊-Scaling Up 3D Human Generation.pdf

1、Scaling UP 3D Digital Human GenerationHao ZhuNanjing University2025/04/113D Digital Human Generation3D Textured ModelPortrait Image“A middle-aged man in a green long-sleeved shirt and blue jeans”Descriptive TextScore Distillation Sampling is slow and unstable.The number of 3D human models is limited

2、.100k Synthesized Data 3D Human Data ScarcityScaling Law+3D Human Generation =？Champ-3D-Driven Human Image Animation ModelReferenceImageDrivingMotionAnimationResultsReferenceImageDrivingMotionAnimationResults Champ-3D-Driven Human Image Animation ModelReference ImageMulti-View PosesResults Multi-Vie

3、w Champ Extend for Better Multi-View Consistency(vi)Output:Multi-view ImagesMVChampc.Image-to-VideogenerationReal-life Images(DeepFashion)(ii)Image sources:Real&Generated imagesFront view,full-body pose of aAge Body Shape CountrySex wearing Clothing andvisible hands.She/He standsagainst a white back

4、ground,evenly lit.a.Attribute CollectionsUniform selectAge:“2030”,.,“7080”Body Shape:“Lean”,.,“Stocky”Country:“Canada”,.,”Japan”Clothing:“Dresses”,.,“Kimono”Sex:“female”,“male”Generated ImagesFLUXb.Text-to-Image generation(i)Text Sources(iii)Condition:Multi-View Pose Sequences360 degree rotation mot

5、ions X 8w times+SMPL-X Fitting&ReposeDiverse animatable motions X 2w times Multi-View Champ Customize Champ For Better Multi-View ConsistencyFeature Fusion 0.5s3D Recovering 0.01sRendering 0.1sGiven ImageGaussian Attribute MapsParametricTemplateGaussian AvatarReshapeLearnableUV TokenUVDecoderUV-Alig

6、nmentTransformerEncoderRenderRendered Posed AvatarAnimateSMPL-X IDOL Instant Photorealistic 3D Human Creation from a Single Image IDOL Results IDOL ResultsReference ImageResultsDriven VideoAccurate Finger Control IDOL Animatable Model IDOL