Single-image 3D face reconstruction, trained without 3D labels

What it does RingNet takes a loosely cropped face image and regresses a complete 3D head mesh using the FLAME morphable model. This repository provides only inference code and pretrained weights from the CVPR 2019 paper. Beyond raw geometry, it can project the input image onto the mesh as texture, or strip out pose and expression to produce a neutral template.

The interesting bit The model learns 3D shape and expression from ordinary 2D photographs—no 3D ground truth was used during training. For practitioners, the neat trick is that inferred FLAME parameters can be neutralized into animation-ready templates for voice-driven pipelines like VOCA.

Key highlights

• Inference-only release: drop in a face crop and get .obj meshes, renderings, or textured outputs
• Exports FLAME shape parameters and camera estimates as .npy arrays for downstream editing
• Can neutralize expression and pose to create templates compatible with VOCA
• Code and pretrained weights are now under MIT license (NoW dataset excluded)
• Companion NoW benchmark evaluation code and validation-set scans live in separate repos

Caveats

• Anchored to Python 2.7, TensorFlow 1.12, CUDA 9.0, and cuDNN 7.3—a stack that is effectively archaeological
• Inference only; the training pipeline and its 3D-supervision-free losses are not included
• Expects a loose face crop as input, so you will need your own face detector for raw images

Verdict Useful if you need FLAME-compatible 3D face meshes for research or animation, but avoid it if you need training code or anything resembling a modern Python environment.