Instax3D: Creating 3D Portrait from a single-view image in Minutes
Keywords: 3D portrait, single-view reconstruction
TL;DR: We propose Instax3D, a generative Gaussian Splatting model with video diffusion prior for fast 3D portrait creation.
Abstract: We study single-view 3D portrait creation, specifically producing a full-head 3D
portrait from a single headshot. This problem faces two challenges: 1) the 2D
image-based personalization methods lack comprehensive 3D awareness due to
the scarcity of multi-view 2D images or 3D assets in the training data, and 2) the
score distillation sampling optimization methods usually take hours to produce a
single 3D asset, making the process quite time-consuming. To overcome these
limitations, we propose Instax3D, a generative Gaussian Splatting model with a
video diffusion prior for rapid 3D portrait creation. We formulate the 3D portrait
creation problem as a “generation and construction” process. Specifically, Instax3D
first synthesizes a consecutive video sequence using a finetuned video diffusion
model, capitalizing on inherent diversity and multi-view knowledge from
the massive video data. Subsequently, Instax3D reconstructs the 3D portrait with
a multi-view FLAME-based Gaussian splatting representation from the generated
video frames, structurally guided by an expressive 3D parametric model. Notably,
given a reference headshot image, Instax3D can generate a 3D portrait in just 10
minutes and render it at 40 FPS. This represents a 10× improvement over previous
mainstream optimization-based methods, which can take between one to two
hours.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 812
Loading