7/n Finally, shout out to the great team at Apple MLR that makes this work possible, Anurag Ranjan, Josh Susskind, @itsbautistam.bsky.social!

6/n More examples of sampled images and pointclouds. Note that we use the same training recipe to train generative models for images and 3D point clouds.

5/n ASFT decodes each coordinate-value pair independently, allowing resolution to change during inference. Our model trained at ImageNet-256 can trivially generate at resolution 2k. Similarly, a model trained for image2point with 16k points can generate much denser pointclouds (e.g., 128k points)

4/n ASFT achieves very strong performance on 3D point cloud generation. On ShapeNet, ASFT outperforms SOTA latent diffusion model LION. On large-scale image2point generation on Objaverse, ASFT also shows very strong results over previous methods like CLAY.

3/n On Imagenet-256, ASFT achieves performance comparable to baselines that apply architectures specifically designed for image generation.

2/n ASFT works directly on ambient space by modeling different data domains as sets of coordinate-value maps. We introduce a conditionally independent point-wise training objective that enables ASFT to make predictions continuously in coordinate space (like in neural fields).