ReTouchGS: Ultralight Render-free Aesthetic Color Enhancement and Transfer for 3D Gaussian Splatting
Organizations: The University of Sydney · Indiana University
Abstract
3D Gaussian Splatting (3DGS) enables photorealistic scene reconstruction. Existing 3DGS color transfer and interactive editing methods are guided by reference images and user-specified color adjustments. However, in many applications, users simply want aesthetically guided color enhancement that automatically makes the scene more visually appealing while maintaining consistent colors and tones across viewpoints. We present ReTouchGS, to our knowledge the first framework for reference-free aesthetic color enhancement of pretrained 3DGS models. It predicts a scene-level color lookup table (LUT) from the model and its associated camera parameters. ReTouchGS also supports reference-guided color transfer by fitting a scene-level LUT using one edited view of the same scene and its corresponding original render. A shared blend-aware color refinement network with only 44.7K parameters refines an analytic color update to realize either color mapping without rendering. It preserves geometry and delivers standard 3DGS assets in seconds. We also design the ReTouch3D benchmark to support the evaluation of aesthetic color enhancement and reference-guided color transfer across viewpoints. Experiments demonstrate high color accuracy and strong generalization to unseen scenes in both aesthetic color enhancement and reference-guided color transfer.
Figures & tables
| Aesthetic Color Enhancement | Reference-Guided Color Transfer | ||||||||||||
| Method | Venue / Year | PSNR | SSIM | LPIPS | Time (s) | Params. | PSNR | SSIM | LPIPS | Time (s) | Params. | ||
| Per-scene Solving and Optimization | |||||||||||||
| Instant Colorization [ 16 ] | CVPR Findings, 2026 | 23.50 | 13.80 | 0.5784 | 0.4911 | 10.95 | 0.9936M | 23.35 | 13.79 | 0.5793 | 0.4883 | 8.98 | 0 |
| BilaRF [ 24 ] | SIGGRAPH, 2024 | 6.30 | 26.97 | 0.9350 | 0.0679 | 122.54 | 0.9940M | 5.73 | 27.87 | 0.9360 | 0.0655 | 121.32 | 381 |
| SAAD [ 3 ] | arXiv, 2026 | 6.18 | 27.05 | 0.9079 | 0.0847 | 3461.78 | 1.008M | 15.46 | 20.10 | 0.8249 | 0.2021 | 340.57 | 14.56K |
| ReparamGS [ 4 ] | arXiv, 2026 | 11.65 | 19.16 | 0.6933 | 0.2701 | 136.08 | 0.9936M | 11.44 | 19.40 | 0.7023 | 0.2658 | 6.02 | 0 |
| Enhancement | Transfer | |||
| Configuration | PSNR | PSNR | ||
| Color update components | ||||
| w/o blended color context & refinement | 6.417 | 26.66 | 5.299 | 27.76 |
| w/o blend-aware color refinement | 5.911 | 27.37 | 4.649 | 28.40 |
| Refinement mechanisms | ||||
| w/o residual propagation | 5.594 | 27.93 | 4.189 | 29.21 |
| Enhancement | Transfer | |||||||
|---|---|---|---|---|---|---|---|---|
| Method | PSNR | SSIM | LPIPS | PSNR | SSIM | LPIPS | ||
| Instant Colorization [ 16 ] | 23.89 | 13.61 | 0.5963 | 0.4920 | 23.62 | 13.66 | 0.5969 | 0.4915 |
| BilaRF [ 24 ] | 6.03 | 26.60 | 0.9521 | 0.0566 | 4.52 | 29.21 | 0.9598 | 0.0521 |
| ReparamGS [ 4 ] | 10.78 | 19.25 | 0.7147 | 0.2476 | 10.06 | 19.58 | 0.7066 | 0.2528 |
| Omni-3DEdit [ 6 ] | 13.56 | 17.84 | 0.6304 | 0.2967 | 13.51 | 17.85 | 0.6305 | 0.2963 |
| ReTouchGS | 5.13 | 27.47 | 0.9598 | 0.0463 | 3.08 | 31.30 | 0.9661 | 0.0421 |
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
| Scene source | Test scenes | Gaussians (millions) | Reconstruction viewpoints | Edit magnitude ( ) |
|---|---|---|---|---|
| DL3DV | 28 | |||
| Mip-NeRF 360 | 9 |
| Scene source | Scene | Gaussians | Viewpoints | Render size | Edit magnitude ( ) |
| DL3DV | 971fb09235c6575e | 1,569,484 | 337 | 10.37 | |
| DL3DV | 97b769bc525dbc21 | 1,249,761 | 421 | 9.10 | |
| DL3DV | 97c7586ee4b3c471 | 621,801 | 403 | 9.03 | |
| DL3DV | 97c8843407a4fef1 | 2,033,317 | 316 | 6.05 | |
| DL3DV | 97cb739ab732aba6 | 1,357,464 | 317 | 15.20 | |
| DL3DV | 98ff6a357347abc2 | 2,388,824 | 337 | 13.70 |
| DL3DV scenes | Mip-NeRF 360 scenes | |||||||
| Method | PSNR | SSIM | LPIPS | PSNR | SSIM | LPIPS | ||
| Aesthetic Color Enhancement | ||||||||
| Original asset | 22.98 | 0.8958 | 0.0926 | 19.85 | 0.8412 | 0.1190 | ||
| Instant Colorization | 14.04 | 0.6120 | 0.4694 | 13.03 | 0.4741 | 0.5589 | ||
| BilaRF | 27.42 | 0.9402 | 0.0627 | 25.59 | 0.9189 | 0.0843 | ||
| SAAD | 27.52 | 0.9131 | 0.0818 | 25.59 | 0.8918 | 0.0936 | ||
| DL3DV scenes | Mip-NeRF 360 scenes | |||||||
|---|---|---|---|---|---|---|---|---|
| Enhancement | Transfer | Enhancement | Transfer | |||||
| Method | Gap | Temporal | Gap | Temporal | Gap | Temporal | Gap | Temporal |
| BilaRF | 0.71 | 0.72 | 2.20 | 1.95 | ||||
| SAAD | 0.90 | 1.24 | 3.34 | 4.33 | ||||
| ReparamGS | 2.20 | 2.13 | 7.05 | 7.15 | ||||
| VIRGi | 1.59 | 1.59 | 5.39 | 5.37 | ||||
| Operator | Asset error | Operator fidelity error | ||||||
|---|---|---|---|---|---|---|---|---|
| Scene source | Operator source | error | Direct | Analytic | Refined | Direct | Analytic | Refined |
| DL3DV | Training mean | 4.89 | 7.01 | 6.52 | 6.23 | 4.51 | 3.82 | 3.30 |
| Aesthetic enhancement network | 2.99 | 6.04 | 5.48 | 5.01 | 4.94 | 4.24 | 3.66 | |
| Image predictor | 2.10 | 5.51 | 4.94 | 4.41 | 5.01 | 4.29 | 3.68 | |
| Enhancer | 2.46 | 5.69 | 5.39 | 4.84 | 5.12 | 4.70 | 4.05 | |
| Enhancer | 1.49 | 5.35 | 5.00 | 4.36 | 5.27 | 4.91 | 4.21 | |
| Scene source | Operator | Color update | PSNR | SSIM | LPIPS | Fidelity error | |
|---|---|---|---|---|---|---|---|
| DL3DV | Grid LUT | Analytic | 4.679 | 28.41 | 0.9145 | 0.0613 | 4.518 |
| Grid LUT | Refined | 4.049 | 29.56 | 0.9229 | 0.0541 | 3.901 | |
| GLUT | Analytic | 4.406 | 28.90 | 0.9203 | 0.0542 | 4.230 | |
| GLUT | Refined | 3.851 | 29.93 | 0.9264 | 0.0493 | 3.669 | |
| Mip-NeRF 360 | Grid LUT | Analytic | 4.556 | 28.37 | 0.9071 | 0.0576 | 4.262 |
| Grid LUT | Refined | 3.990 | 29.26 | 0.9193 | 0.0491 | 3.684 |
| Enhancement | Transfer | |||
|---|---|---|---|---|
| Operator + update | PSNR | PSNR | ||
| VCG + analytic | 8.388 | 24.95 | 8.322 | 25.03 |
| Grid LUT + analytic | 5.911 | 27.37 | 4.649 | 28.40 |
| Grid LUT + refined | 5.502 | 28.08 | 4.035 | 29.49 |
| NamedCurves MIT5K-DPE | Image-adaptive 3D LUT | ||||
|---|---|---|---|---|---|
| Predictor input | Training scenes | Validation | Held-out views | Validation | Held-out views |
| Original renders | 36 | 3.525 | 3.439 | 2.201 | 2.573 |
| Original renders | 72 | 2.858 | 2.661 | 1.665 | 1.890 |
| Original renders | 145 | 2.464 | 2.327 | 1.298 | 1.656 |
| Original renders | 189 | 2.324 | 2.124 | 1.292 | 1.561 |
| Projected CNN | 145 | 3.522 | 3.203 | 1.830 | 1.881 |
| DL3DV | Mip-NeRF 360 | |||
| Configuration | Enhancement | Transfer | Enhancement | Transfer |
| Color update components | ||||
| w/o blended color context & refinement | 6.043 | 5.264 | 7.580 | 5.406 |
| w/o blend-aware color refinement | 5.477 | 4.679 | 7.261 | 4.556 |
| Refinement mechanisms | ||||
| w/o residual propagation | 5.129 | 4.179 | 7.038 | 4.220 |
| Scene-consistent target | Per-view target | Unobserved regions | ||||||
|---|---|---|---|---|---|---|---|---|
| Scene source | Task | Analytic | Refined | Analytic | Refined | Direct | Analytic | Refined |
| DL3DV | Enhancement | 5.48 | 5.01 | 6.34 | 5.90 | 6.662 | 6.564 | 6.301 |
| Transfer | 4.68 | 4.05 | 5.76 | 5.23 | 5.992 | 6.173 | 5.530 | |
| Mip-NeRF 360 | Enhancement | 7.26 | 7.04 | 9.33 | 9.19 | 7.521 | 7.480 | 7.045 |
| Transfer | 4.56 | 3.99 | 7.31 | 6.98 | 5.383 | 5.309 | 4.623 | |
| Aesthetic Color Enhancement | Reference-Guided Color Transfer | |||||||
| Method | Time (s) | Network (M) | Edit vars. | Time (s) | Network (M) | Edit vars. | Prep. vars. | GPU |
| Instant Colorization | 10.95 | 0.9936 | 82.24M | 8.98 | 0 | 82.24M | 0 | RTX 5090 |
| BilaRF | 122.54 | 0.994 | 381 | 121.32 | 0.000381 | 381 | 0 | RTX 5090 |
| SAAD | 3,461.78 | 1.008 | 14.56K | 340.57 | 0.01456 | 14.56K | 0 | RTX 5090 |
| ReparamGS | 136.08 | 0.9936 | 612 | 6.02 | 0 | 612 | 164.5M | RTX 5090 |
| FPGS | 633.82 | 31.3 | 0 | 17.41 | 30.3 | 0 | 119.1M | RTX 5090 |
3D Scene Editing
3D scene editing studies how to make controlled changes to objects, appearance, layout, and motion while keeping a scene coherent across views.
See all 94 papers