Wonder3D: Single Image to 3D using Cross-Domain Diffusion paper page: huggingface.co/papers/2310.15… introduce Wonder3D, a novel method for efficiently generating high-fidelity textured meshes from single-view images.Recent methods based on Score Distillation Sampling (SDS) have shown

It's always important to know the limitations of large models, especially in applying them to robotics/AD. I believe the lack of 3D spatial reasoning will be tackled by the scalable 3D training with posed embedded 2D images.

Jerome Revaud I thought you must be joking so I ran it on these two of my own pictures (excuse my dirty dishes 😉) and wow, it works beautifully.

Excited to introduce 𝗗𝗦𝗜𝗡𝗘! (#CVPR2024) baegwangbin.github.io/DSINE/ We push the limits of single-image surface normal estimation by rethinking the inductive biases needed for the task. See you in Seattle!

Except for the multiview MAE-style transformer's pre-training, the pointmap part is trained on 8.5Mx2=17M imgs with 3D GT points (8 datasets, ScanNet++, CO3D-v2, MegaDepth, ARK-it Scenes, etc.), which is 2x bigger than Objaverse Zero123's 9.6M imgs. Is this scale enough for 3DV?

Play SoTA Gaussian Splatting SLAM system with interactive GUI is enjoyable! Congrats on the code release!

Best 3D Object Gen tool I have tried so far. The multi-view input requires a paid subscription tho. Very interested in the tech details and the roadmap. Deemos Tech The scale of 3D mesh data (img2mesh) seems enough for object-level. How about scene-level? Multiview/Seq-img2img?

This is an opportunity to do a PhD with me at Imperial College, fully funded and starting in October this year. Apply via the link below by 12th June next week. On-sensor vision will be very important to the future of low power vision in robotics + AR/VR. jobs.ac.uk/job/DHT079/res…

👍SD3 -> diffusers ✅Better text generation in images and varying aspect ratios 🔍Rectified Flow Matching +DiT+Separate Image/Text Encoder

🥵#CVPR2024

Text2Video(16 frames), camera condition with plucker embedding + learnable FF, finetune from snap video model (36 × 64 resolution), FIT backbone, 1dx8xA100-40GB training on 65K video clips from Re10K.