Framestore has published a 54-minute breakdown of the 4D Gaussian splatting pipeline behind the Kryptonian holograms in Superman.
Kevin Sears, a CG supervisor at Framestore Montreal, delivered the talk as a webinar on September 22, and Framestore posted it to YouTube on October 21. The holographic messages from Superman's Kryptonian parents were splat performances reshaped in Houdini, and the look ended up in about 30 shots. The shoot took place in August 2024 in Ipswich after Framestore persuaded the client not to move the setup to Los Angeles, with James Gunn supervising remotely.
Infinite Realities captured the performance on 192 cameras recording 12 megapixel images at 24 fps, and each frame resolved to about 1.6 million splats. The capture volume is roughly the size of a small bedroom, and Sears said it was the first time Infinite Realities had put two performers that close together in it. The client chose a single take of about 88 seconds, which was cut up to cover every shot. That works out to about 2,000 frames per camera and a little over 400,000 images.
Infinite Realities trained the sequence in Postshot on 20 processing nodes with RTX 4090 GPUs, and the final splat took several days after a few rounds of rising quality. Early results chattered around faces, hair and the white costumes, and a longer retrain of several more hours per frame cleaned them up. The source images were 8 bit and the data came to about 5 TB per take. Infinite Realities delivered a PLY sequence with the actors cropped in 3D and the ground plane removed.
At the time, Postshot could not load animated cameras, Houdini did not load splats natively, Nuke had no splat plugin and there was no relighting. ShotGrid fields from editorial drove a Houdini setup that wrote a unique PLY sequence per shot, while layout placed an unanimated proxy model at scales from about 10 feet to about 6 inches. The splats could not be moved inside the renderer, so assistant technical directors built a constraint system that baked one camera per splat cluster, so a shot with three or four clusters had three or four render cameras.
Rendering ran through Inria's SIBR Viewer, with opacity and alpha controls added by Infinite Realities and rotations and time offsets added later. Sears rendered every splat himself through a command line batch, pushing alpha above the solved value so the front of a face never revealed the back of the head, and output 4K PNG frames. The FX team ran a newer Houdini than the pipeline supported, used CG Nomads' GSOPs to write splats out natively, and explored particles shedding from the edges before abandoning them.
A native Nuke splat plugin arrived late in the project and was tested, but the final shots did not use splats inside Nuke. Compositing applied the edge opacity treatment in 2D with little use of depth channels, along with heavy grading and denoising. Every frame was trained independently with no temporal coherence, so temporal denoising in comp reduced the chatter along the edges. Asked about a fallback, Sears said there was not really one beyond projecting the 192 camera feeds onto cards, which was never needed.
Sears said Houdini and Nuke now load splats natively and that he would not necessarily reuse this workflow today. According to Sears, Framestore has used splats for environment work on commercials, but he has not seen the studio use them to replace animated characters. The full talk is available on Framestore's YouTube channel.
