Learning Path
Capture & Training Workflow
From lighting and lenses to SfM and training params: an end-to-end pipeline from reality to Gaussians
Level
Intermediate
Duration
约 4 小时
Audience
Creators, photographers and developers who want to capture and train splats themselves
Prerequisites
- Completed Path 01 or already grasp the basics of 3DGS
- Have a phone, camera or action cam available
- Willing to follow capture protocols and tolerate training waits
After this path you can
- Build a pre-shoot checklist for 3DGS-ready captures
- Master capture routes for small rooms, objects and large spaces
- Choose between cloud and local training with cost awareness
- Use SuperSplat and splat-transform for QC, cleanup and export
- Establish a reusable capture-preprocess-train-export workflow template
Modules
8 modules, layered understanding
- Module 01Article25 分钟
Pre-shoot Engineering: Light, Sharpness & Overlap
Treat capture as engineering rather than improvisation. This module consolidates the three foundational variables — lighting, focus sharpness and frame overlap — and offers a pre-shoot self-check list to keep training runs from paying interest on lazy capture.
What's inside
- 01Why 3DGS cares more about light consistency than color temperature
- 02Judging sharpness: from focal points to motion blur
- 03Overlap rule of thumb: 70% horizontal, 60% vertical
- 04The eight-item pre-shoot checklist
- Module 02Tutorial30 分钟
The Three-Layer Capture Method
Picture the space as an onion: shoot in three concentric layers. This module walks through the回字-shaped path design, lens-height hierarchy, and how to keep cameras from photographing each other inside cramped venues.
What's inside
- 01Geometric meaning of three rings: high, mid and low viewpoints
- 02回-shaped paths and loop closure: making SfM converge
- 03Small-space tactics: avoiding lens-in-frame and occlusion
- 04Hands-on: a full route with Scaniverse / Luma
- Module 03Tutorial30 分钟
Object Video Capture: Phone Hands-on
Recording an object on your phone is the lowest-barrier entry into 3DGS. This module shares reusable grip, pacing and shot lists so a single object becomes training-ready footage within ninety seconds.
What's inside
- 01Grip and stabilization: killing high-frequency jitter
- 02Pacing: a 15-degree-per-second orbit
- 03Three-shot script: top-down, eye-level, low-angle
- 04Luma vs Scaniverse: experience-level differences
- Module 04Tutorial30 分钟
Object Photo Capture: Studio Setup
Stepping into a studio means controlling every photon. This module covers the turntable-plus-black-cloth minimal rig, ideal angular spacing between shots, and how KIRI's Remy workflow turns photo pipelines into a near push-button experience.
What's inside
- 01Turntable plus dual softbox: a minimal studio
- 02Angle spacing: every 5 degrees or every 10
- 03Polycam vs KIRI Engine in the photo pipeline
- 04Where Remy's automation breaks and what it costs
Related Terms
- Module 05Tutorial35 分钟
Action Cams & 360 Cams: Large-space Capture
When the venue grows from a single room to an entire courtyard, phones run out of breath. This module benchmarks Insta360 panoramic rigs, DJI Terra aerial routes, and Monogram's Japan field practice so you can pick a kit matching your terrain.
What's inside
- 01360 cam advantages of FOV redundancy and dewarping pitfalls
- 02Drone surveying: DJI Terra waypoint planning
- 03Action cam choreography: handheld stability and trajectory
- 04After the shoot: frame extraction and redundancy trimming
- Module 06Video35 分钟
Local Training 101: Postshot / Brush / Lichtfeld Studio
Local training marks the crossover from hobbyist to practitioner. Anchored on three desktop tools that are free or low-cost, this module spells out VRAM budgets, step counts and convergence heuristics — including a what-to-do-when-it-crashes routine.
What's inside
- 01Postshot: drag-and-train desktop experience
- 02Brush: customizable open-source trainer
- 03Lichtfeld Studio: a research-leaning engineered trainer
- 04Reading training logs: loss curves and splat-count evolution
Related Terms
- Module 07Article25 分钟
Cloud Training: Luma / Zhitianxia / Pointcosm
When local hardware buckles, the cloud is the other leg. This module benchmarks Luma's minimalist uploads, Zhitianxia's enterprise-leaning platform, and Pointcosm's package-based pricing, with a selection matrix indexed by scene scale.
What's inside
- 01Upload UX: from front-end UI to failure retries
- 02Benchmarking training time against price
- 03Export formats: ply vs splat vs spz trade-offs
- 04Enterprise platforms vs packaged plans
Related Terms
- Module 08Tutorial30 分钟
Validation & Export: QC + Format Choice
Training done is just round one of quality control. This module shows how to delete floaters in SuperSplat, convert formats with splat-transform, and grade outputs against a simple PSNR/SSIM/LPIPS rubric.
What's inside
- 01Lasso and delete floaters in SuperSplat
- 02Convert ply / splat / spz with splat-transform
- 03Self-evaluation with PSNR / SSIM / LPIPS
- 04Final delivery checklist before handing off
Related Terms
Put it to work
After this path, head to the tool index, assemble a capture-plus-trainer kit, and scan the room you know best as your first homework.
Browse Tools Index