Chapter 7 — Built-in Quality Presets

WHERE

Inspector → Presets section → group "Classic" → entry "Quick". UUID suffix …001.

TECHNICAL

Diagnostic preset with 1,000 iterations, classic (adaptive) densification strategy, and a training resolution scale of 0.25× (input images are shrunk to 25 % before training). Not intended for delivering a finished scene, but to quickly determine whether the setup (camera poses, point cloud, image sequence) shows any meaningful movement in the loss values at all. On an M3 Ultra typically under 30 seconds on 50–200 images. The small resolution obscures true image quality but keeps memory demand and render cost very low. Also chosen automatically as the default on first launch if the system has less than 10 GB of RAM.

WHERE

Inspector → Presets section → group "Classic" → entry "Preview". UUID suffix …002.

TECHNICAL

5,000 iterations of Classic densification, 0.5× resolution scale, double the learning rates compared to the default. Densification (cloning + splitting) is active over the first 2,500 iterations, after which only pruning continues. Default preset for systems with ≥ 10 GB RAM. On an M3 Ultra typically 90 seconds to 3 minutes for a 200-image scene. Delivers a usable impression of the geometry and camera pose, but textures are visibly blurred — the 0.5× render resolution can't be bypassed simply by retraining with P3 or P4 afterwards, because the learning rates are calibrated for the halved resolution.

WHERE

Inspector → Presets section → group "Classic" → entry "Balanced". UUID suffix …005.

TECHNICAL

20,000 iterations of Classic densification at full image resolution. Densification runs over the first 15,000 iterations, with a densify interval of 100 starting at iteration 3,000. In this range, returns clearly flatten out: quality stabilises between iteration 18,000 and 22,000, and only the jump to Quality (P4) brings noticeably more again. On an M3 Ultra typically 30–60 seconds on 200 images, 5–8 minutes on 1,000+ images.

WHERE

Inspector → Presets section → group "Classic" → entry "Quality". UUID suffix …003.

TECHNICAL

35,000 iterations of Classic densification with "Opacity Decay" (HTGS, Eurographics 2025): after each densify cycle, the opacity of all existing Gaussians is multiplied by a factor < 1.0, which reliably removes Gaussians that have become inactive during pruning — a visibly cleaner result than without this step at the same iteration count. SSIM loss is enabled (ssimWeight=0.05). On an M3 Ultra typically 2–4 minutes on 200 images. On object scenes, the strongest of the classic variants. Note: needs ~3–5 GB of GPU memory; on 8 GB systems, P3 is the safer choice.

WHERE

Inspector → Presets section → group "Classic" → entry "Ultra Detail". UUID suffix …008.

TECHNICAL

Around 35,000 iterations of Classic densification — the sharpest recipe in the Classic group. At a comparable compute time, it outperforms the MCMC "Quality" preset (P8), making it the strongest quality preset in the Classic group and the sharpest Classic variant RadianceKit ships. On an M3 Ultra typically in the same time frame as P4 Quality (2–5 minutes on 200 images), but requires somewhat more GPU memory; on 8 GB systems, P3 remains the safer choice.

WHERE

Inspector → Presets section → group "MCMC" → entry "Preview". UUID suffix …006.

TECHNICAL

60,000 iterations of MCMC densification (3DGS-MCMC, NeurIPS 2024) with a cap of 100,000 Gaussians. MCMC replaces the heuristic clone/split logic with Markov-Chain-Monte-Carlo relocation: dead Gaussians are re-placed via sigmoid-weighted sampling depths, giving a controlled and reproducible number of Gaussians. The cap hard-limits the maximum count at 100K — saving memory and render time, at the cost of detail. On an M3 Ultra typically 5–8 minutes on 200 images. Serves as an "MCMC functionality test" — helps judge whether switching from Classic to MCMC would be worthwhile before you invest more time in P7 or P8.

WHERE

Inspector → Presets section → group "MCMC" → entry "Balanced". UUID suffix …007.

TECHNICAL

120,000 iterations of MCMC with a cap of 150,000 Gaussians. The middle MCMC tier — nearly the final Gaussian count of P8 Quality, but only 60 % of the iterations. The result therefore lags a bit behind P8 Quality but costs only about half the waiting time. On an M3 Ultra typically 8–15 minutes on 200 images. Uses a method that scales the effective Gaussian cap to the point density of the input SfM point cloud (see T75 in Chapter 6).

WHERE

Inspector → Presets section → group "MCMC" → entry "Quality". UUID suffix …004.

TECHNICAL

200,000 iterations of MCMC with a cap of 150,000 Gaussians, SSIM loss 0.05, MCMC noise decay over 80 % of the iterations. Image quality is on par with the best Classic variant (P4), but MCMC gets there with about 71 % fewer Gaussians (150K vs ~524K) — decisive if you want to deliver the result on the web, because the smaller cloud produces noticeably smaller export files. Training time on an M3 Ultra is typically 20–35 minutes on 200 images; on 1,000+ image sets more like 1–2 hours. The best choice when maximum image quality at minimum final size is desired.

WHERE

Inspector → Presets section → group "Capture Class" → entry "Drone / Aerial". UUID suffix …010.

TECHNICAL

Capture-Class preset for aerial and drone orbits of buildings and landscapes. MCMC densifier, 40,000 iterations, cap of 1.5 million Gaussians, SSIM loss 0.5 plus edge-aware term 0.1. Decisive is the anisotropy penalty with weight 0.003 at a ratio threshold of 6 — the "spaghetti killer" against the typically needle-shaped artefacts that drone footage produces. Validated on a real DJI 4K drone flight over the Pensford Viaduct (checked by eye, not merely by metrics).

WHERE

Inspector → Presets section → group "Capture Class" → entry "360° Walkaround". UUID suffix …011.

TECHNICAL

Capture-Class preset for 360° walkaround videos. MCMC densifier, 40,000 iterations, cap of 1.5 million Gaussians, SSIM loss 0.5 plus edge-aware term 0.1, gentle anisotropy penalty (weight 0.001 at ratio threshold 15). Person and sky masking are active. The preset expects a 360° equirectangular video, which is internally reprojected into roughly 90°-wide perspective crops before training starts. Validated on 8K 360° walkarounds shot with a selfie stick (Monument scene, checked by eye).

WHERE

Inspector → Presets section → group "Capture Class" → entry "360° Walkaround (High Detail)". UUID suffix …013 (opt-in).

TECHNICAL

Opt-in Capture-Class preset for 360° walkaround videos with maximum detail. Hybrid densifier (classic abs-gradient cloning/splitting + MCMC noise + relocation), 40,000 iterations, cap of 1.5 million Gaussians, anisotropy penalty 0.0015 at ratio threshold 15, SSIM loss 0.2 and edge-aware term 0. On 360° material it beats the standard MCMC preset "360° Walkaround" (P10) on quality metrics and on visible confetti, and does so with about a third of the splat count. Deliberately sits opt-in alongside the standard 360 preset until it has been validated on more scenes.

WHERE

Inspector → Presets section → group "Capture Class" → entry "Photo / Object". UUID suffix …012.

TECHNICAL

Capture-Class preset for object orbits made from sharp individual photos (no video). Hybrid-t1 densifier (with relocation), 40,000 iterations, cap of 1.5 million Gaussians, SSIM loss 0.5 plus edge-aware term 0.1, gentle anisotropy penalty (weight 0.001 at ratio threshold 15), opacity decay 0.9995 every 50 iterations, no masking. Validated on 163 high-resolution 41 MP photos of a skeleton (checked by eye). Few views (up to about 600) stay under the hybrid-collapse threshold.

WHERE

Inspector → Presets section → group "Capture Class" → entry "Photoset / Unordered".

TECHNICAL

Capture-Class preset for an unsorted photoset — shots with no known order, for instance from multiple sessions or several people. 40,000 iterations. For this capture type the app always aligns camera positions using the native method, regardless of the otherwise configured setting (see Chapter 9).

WHERE

Inspector → Presets section → group "Capture Class" → entry "Maximum Quality Native".

TECHNICAL

The most expensive recipe on the list: 80,000 iterations with a Gaussian ceiling of 20 million. It sets both the training and the highest native alignment tier in one step (recipe E3, see Chapter 9) — the direct route to the highest quality of the native path, without setting the individual values by hand. Correspondingly long runtime and correspondingly high memory demand: on weaker systems, one of the other Capture-Class presets is the safer choice.

WHERE

Inspector → Presets section → group "Hybrid" → entry "Balanced (Hybrid)". UUID suffix …009.

TECHNICAL

20,000 iterations with the hybrid densification strategy: classic gradient-driven cloning/splitting places capacity where the loss needs it, MCMC SGLD noise keeps exploring, and dead Gaussians are relocated instead of being lost during pruning. Opacity decay replaces opacity resets; an anisotropy penalty (weight 0.001, ratio threshold 15) keeps needle-shaped splats in check. The Gaussian cap scales with the scene (150K base, scene-aware ×3.0). Compared to pure MCMC at the same budget, this delivers somewhat more image quality while needing 20–30 % fewer Gaussians. On an M3 Ultra typically 5–10 minutes on 200 images.