... CLIPScore und PickScore für Prompt-Treue sowie referenzbasierten wie Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity (SSIM) und Learned Perceptual Image Patch Similarity (LPIPS). 2/4
Apple publicó SHARP: subes una foto y en menos de 1 segundo te devuelve una escena 3D navegable. Redujo LPIPS 25-34% y DISTS 21-43% frente al mejor modelo anterior. Es el paper código abierto de Apple. github.com/apple/ml-sharp
ViewMask-1-to-3 outperforms the baseline on both GSO and 3D-FUTURE:
🏆#1 average ranking across image-level metrics (PSNR, SSIM, LPIPS, CD, IoU)
🏆Up to 10.6% IoU higher on 3D-FUTURE than continuous diffusion models
🏆Ranked top tier on GenEval
If we assume that ALL textures had zeroed-out metallic channel, the results would look something like this:
SSIM:
BC7ENC: 0.93925
MY ENC: 0.93565
RMSE:
BC7ENC: 5.1213
MY ENC: 5.3502
LPIPS:
BC7ENC: 0.05145
MY ENC: 0.0456
And if I use 6x6 principal axis, it loses even on LPIPS. I guess 6x6 PA degrades small detail too much.
Also, note that most textures have a completely black metallic channel, so results are slightly skewed.
Downloaded 763 1k ARM textures from polyhaven and tested b7enc and my encoder (with 4x4 principal axis) using SSIM, RMSE and LPIPS
I used L5 bc7enc mode as L18 is too slow and L5 seems to be closer to my encoder under the hood
So my encoder seems to score better only on LPIPS...
RT-Splatting is a breakthrough for real-time rendering of glass, plastic, and other semi-transparent surfaces. For the first time, you get both crisp reflections and clear see-through transmission—no more choosing between blurry mirrors or hidden backgrounds.
The secret: each Gaussian encodes its geometry (does a ray hit?) and optics (how much light is absorbed?) separately, so reflections and transmission are disentangled. Their “Specular-Aware Gradient Gating” keeps artifacts in check, while hybrid surface/volume rendering brings photorealism at 33 fps—even in complex scenes.
On 8 tough real-world datasets, RT-Splatting outperforms all prior Gaussian methods by 0.35–2.0 dB PSNR and up to 20% better LPIPS, with the biggest gains on transparent regions.
Scene edits are easy: tweak glass tint or reflection strength after training by scaling a few parameters—no retraining needed.
This is a huge leap for AR, digital twins, automotive, and photorealistic 3D product previews.
Get the full analysis here: yesnoerror.com/abs/2605.1826…
// alpha identified
// $YNE