What's Changed

hot fix

Full Changelog: v1.1.0...v1.1.1

Highlight

Multiple densification strategies are now supported in gsplat, including

• DefaultStrategy(): The original 3DGS densification process.
• DefaultStrategy(absgrad=True): Densification process from "AbsGS: Recovering Fine Details for 3D Gaussian Splatting"
• MCMCStrategy(). Densification process from "3D Gaussian Splatting as Markov Chain Monte Carlo"

What's Changed

• Update 3DGS examples to use the latest nerfview from pypi by @hangg7 in #206
• docs: update README.md by @eltociear in #210
• Added a guidance for installing gsplat on Windows by @KevinXu02 in #213
• Minor fix for _QuatScaleToCovarPreci back with sparse grad by @liruilong940607 in #214
• Fix wrong description in rasterization API doc by @yifanlu0227 in #216
• support 2^n+1 channels for rasterize_to_pixels by @liruilong940607 in #218
• workaround for n-dim rendering with n>32 by @wuzirui in #217
• Minor fix for _RasterizeToPixels back to avoid NaNs by @opipari in #235
• clean up the shape of the colors by @liruilong940607 in #228
• static shared mem -> dynamic shared mem by @liruilong940607 in #229
• updating background for different rendering modes [D, ED, RGB+D, RGB+… by @vye16 in https://github.com//pull/240
• Fix max_range in _torch_impl._rasterize_to_pixels() by @liruilong940607 in #242
• Implement 3DGS-MCMC in gsplat. by @jefequien in #238
• sqrt int to float by @liruilong940607 in #246
• Clamp ratios in MCMC's compute_relocation by @yzslab in #247
• add poetry files, and fix numpy error and no data error. by @MasahiroOgawa in #251
• Separate C source into many translation units by @fwilliams in #255
• Fwilliams/fix headers and ninja by @fwilliams in #256
• template everything but don't actually support multiple precisions yet by @fwilliams in #259
• fix glm path in JIT building by @liruilong940607 in #260
• Fwilliams/atomics by @fwilliams in #261
• set the MAX_JOBS to 10 if not set. by @liruilong940607 in #262
• float3 -> vec3; float2 -> vec2; by @liruilong940607 in #267
• Start adding support for double, float16, and bfloat16 by @fwilliams in #270
• fix test_fully_fused_projection_packed test by @liruilong940607 in #289
• Add optional masks in rasterize_to_pixels() to support Grendel by @alexis-mmm in #284
• Abstracting out the GS heuristics into a Strategy class by @liruilong940607 in #278
• Bump to 1.1.0 by @liruilong940607 in #290

New Contributors

• @yifanlu0227 made their first contribution in #216
• @wuzirui made their first contribution in #217
• @opipari made their first contribution in #235
• @jefequien made their first contribution in #238
• @yzslab made their first contribution in #247
• @MasahiroOgawa made their first contribution in #251
• @fwilliams made their first contribution in #255
• @alexis-mmm made their first contribution in #284

Full Changelog: v1.0.0...v1.1.0

A Major Update to V1.0.0

• Comparing to the official implementation, gsplat enables up to 4x less training memory footprint, and up to 2x less training time on Mip-NeRF 360 captures, and potential more on larger scenes.

• Support extremely large scene rendering, which is magnitudes faster than the official CUDA backend diff-gaussian-rasterization.

• Extra features, including batch rasterization, N-D feature rendering (faster), depth rendering, sparse gradient etc.

What's Changed

• Add backward gradient for background by @KevinXu02 in #170
• enable doc versioning by @liruilong940607 in #185
• Fix version value for Sphinx by @brentyi in #191
• Improve numerical stability of _torch_impl.project_cov3d_ewa by @oseiskar in #182
• Update project_pix to use eps by @SimonF24 in #192
• create JIT compilation configs if necessary by @ychfan in #164

New Contributors

• @KevinXu02 made their first contribution in #170
• @SimonF24 made their first contribution in #192
• @ychfan made their first contribution in #164

Full Changelog: v0.1.11...v.0.1.12

What's Changed

• Fast SH implementation by @jb-ye in #165
• return result array for lower sh degrees by @YertleTurtleGit in #171
• Implement AbsGS by @liruilong940607 in #166

New Contributors

• @YertleTurtleGit made their first contribution in #171

Full Changelog: v0.1.10...v0.1.11

What's Changed

• Fix missing +0.5 in calculating uv coordinate by @jb-ye in #151
• Check for normalized quats. by @Mxbonn in #155

New Contributors

• @Mxbonn made their first contribution in #155

Full Changelog: v0.1.9...v0.1.10

What's Changed

• Small CUDA changes for nd rasterization by @kerrj in #144
• Remove redundant projection matrix by @oseiskar in #97
• Approximate view matrix gradient for pose optimization by @oseiskar in #127

New Contributors

• @oseiskar made their first contribution in #97

Full Changelog: v0.1.8...v0.1.9

What's Changed

Some quality improvements because of:

• Fix v_mean3d in project_g and v_conic calculation in *rasterize_backward by @jb-ye in #139
• add backprop grad of opacity compensation by @jb-ye in #140

Full Changelog: v0.1.7...v0.1.8

What's Changed

• Eval table by @AdamRashid96 in #134
• add device guard by @liruilong940607 in #135
• Fix backprop grad of cov2d / scales and unit tests by @jb-ye in #136

New Contributors

• @AdamRashid96 made their first contribution in #134

Full Changelog: v0.1.6...v0.1.7

What's new

The python interface for tile_bounds has been abstracted away, and instead project and rasterize take in a block_width parameter controlling the tile size as well as the CUDA block size. CUDA was refactored to allow dynamic block sizes, which enables ND rasterization speedups by taking better advantage of shared memory.