BSD Studio User Manual

Download · Sign-in & license · Data · 3D reconstruction · Delivery · Purchasing & support

Figure 1 The current BSD Studio interface; details may change between versions

Applies to: BSD Studio v10.5.0 | November 2026

About this manual: This is the official BSD Studio technical and operating guide, published by Zhongxin Lianzhi (Wuhan) Information Technology Co., Ltd. It covers download and installation, accounts and licensing, data preparation, alignment, Gaussian training, viewing and editing, project management, purchasing and support.

# How to read

This manual is for photogrammetry and 3D reconstruction users new to BSD Studio, and can also serve project leads as a reference for purchasing, deployment and troubleshooting. Read chapters 1–4 to build your first model; if you already have external alignment results, go straight to chapter 6.

Goal	Read
Understand the product in ten minutes	Chapters 1–3: workflow, hardware, installation and license
Build a model from samples or your photos	Chapters 4–8: data, automatic workflow, alignment and training
Deliver results	Chapters 9–10: viewing, editing, export and project management
Purchasing, maintenance and troubleshooting	Chapters 11–13: purchasing, support and FAQ

Glossary

Term	Meaning
Alignment / SfM	Solving camera positions and orientations from overlapping photos and generating a sparse point cloud.
Gaussians / 3DGS	A scene represented by many 3D Gaussians with color, scale and opacity that can be viewed in real time.
Training images	The final images used for Gaussian training; they may be original photos, video frames or undistorted panoramic views.
Mask	Pixel-level masks that exclude regions such as people or sky; not every scene needs them.
PLY / Mesh	PLY is commonly used for Gaussians and point clouds; a mesh is a triangle surface, commonly saved as PLY, OBJ or STL.

Safety note: For real projects, keep separate backups of the source data and alignment results; never keep the only copy in a temporary project folder.

# 1 Product and workflow

BSD Studio is Windows desktop 3D reconstruction software that creates and displays Gaussian scenes from photos, regular video, native panoramic video or external alignment results. A typical project goes through five stages: import, alignment, training, review and delivery.

Figure 2 The standard workflow from input to delivery

1.1 Choose your entry point

Entry point	When to use	What you do
Automatic workflow	New users, sample data, a single well-organized project	Drag in files, a folder or a video, check the settings in the wizard and start.
Step-by-step workflow	You want to check frame extraction, alignment and training one by one	Create a project, then run import, alignment and training in order.
External alignment import	Alignment already done in DJI Terra, RealityScan, Metashape, etc.	Choose the results on the matching import page, then check the images and camera poses.

1.2 Before you start

Make sure you may legally use the data; photos have enough overlap, sharpness and angle coverage.
A stable NVIDIA driver and enough VRAM and disk space; large high-resolution scenes need more headroom.
Use a separate working folder for each project; do not move source data or clear temporary files while training.

# 2 Hardware, download and installation

Item	Minimum	Recommended
Operating system	Windows 10/11 64-bit	Windows 11 64-bit with stable system and GPU driver versions
CPU	64-bit x86, 4 cores / 8 threads or more e.g. Intel Core i5 / i7, 8th–10th gen or AMD Ryzen 1000–2000 series	8 cores / 16 threads, or 12 cores / 24 threads and above e.g. Intel Core i7 / i9, 12th–14th gen or AMD Ryzen 5000 / 7000 / 9000 series
GPU	NVIDIA discrete GPU (GTX 1060 6 GB or better)	NVIDIA RTX 30 / 40 / 50 series, 8–24 GB VRAM or more
Memory	8 GB	16 GB or more; 64 GB or more for large scenes
Storage	A local disk with enough free space	SSD with room for source data, alignment files, checkpoints and exports

2.1 Hardware and system

GPU compatibility: BSD Studio's native CUDA kernels are compiled and tested for all mainstream NVIDIA architectures: Pascal (GTX 1060 / 1070 / 1080 / 1080 Ti, sm_61), Turing (RTX 20 and GTX 16 series), Ampere (RTX 30), Ada Lovelace (RTX 40) and the latest Blackwell series. On older cards such as the GTX 10 series we recommend at least 6 GB VRAM and a stable NVIDIA driver 535.xx or newer.

2.2 Download from the website

Go to https://www.bsd3dgs.com/en/ and click “Download”. Check the version number, installer name and release date, and download the official installer.

Save the Windows installer to a local disk. If your browser or Windows asks about the file's origin, verify the domain and source; do not use third-party mirrors.

Close any older version of BSD Studio before installing and leave room for the app and project data. Choose the install location in the wizard, then start the app from the Start menu or the desktop shortcut.

On first launch, allow network access so your account and license can be verified. If security software blocks the app, check its source before allowing it.

2.3 First launch checks

The main window opens normally, with the license status and language switch in the top-right corner.
Open the Help page to see the version number; include it when you report a problem.

# 3 Sign-in, license and plan limits

3.1 Sign up / sign in with email

The sign-in window opens on first launch, or from the “License” entry in the top-right corner.

Enter your email address and click “Send code”. Check your inbox (and spam folder) and enter the code before it expires. A new account is created automatically the first time you verify an email address.

Click “Sign in” and check the license status in the main window. If it does not refresh, check your network and reopen the license page.

Account security: Only enter sign-in codes in the official app or on the official website. Our support staff will never ask for your code.

3.2 Basic vs. Pro

Capability	Basic	Pro
Final training images	Up to 600 per project	No 600-image limit; limited only by your hardware
Regular video	Up to 10 minutes, and no more than 600 final training images	Depends on license and hardware
Native panoramic video	Up to 1 minute, and no more than 600 final training images	Depends on license and hardware
External alignment import	As shown in the app for your license	Included
Tiled training, model / mesh export	High-precision mesh extraction and tiled large scenes (depending on system resources)	Full Pro feature set, seamless tiled processing of large scenes and high-resolution model export

“600 images” means the final images that go into training, not the number of files in the import list. Placing external alignment results into the project folder by hand does not bypass the image-count check on AT/images before training.

Frame extraction: A 10-minute video sampled at 1 FPS gives about 600 frames; at 2 FPS it gives about 1,200, which exceeds the Basic limit. Undistorting panoramic video often turns one frame into several views, so estimate the final image count first.

# 4 Preparing and capturing data

4.1 Photo projects

Shoot continuously around the object or scene with clear overlap between neighboring photos; do not capture only the front while missing the sides, back and top.
Keep images sharp and exposure stable; avoid heavy motion blur, overexposure, purely reflective surfaces and large textureless areas.
For large indoor scenes, cover doorways, corners and long corridors that connect areas; for large outdoor scenes, capture both close-up detail and overall structure.
Put the images for one project in their own folder. Remove obvious duplicates, out-of-focus and unrelated photos before importing.

4.2 Video projects

Regular video supports common formats such as MP4 and MOV, and frames are extracted at the frame rate you set. On Basic, set 1 FPS explicitly in the import settings and check the final image count after extraction. Long or fast-moving video adds duplicate or blurry frames and does not necessarily improve results.

Native dual-fisheye panoramic video (.insv / .osv) is supported; check that it is detected as panoramic when importing. The app extracts frames and then undistorts them into training views. The number of views per frame depends on the camera and undistortion settings, so one minute does not automatically mean 60 training images.

4.3 Pre-capture checklist

Check	Pass criteria
Coverage	All main visible sides of the subject are captured, with no obvious gaps.
Overlap	Neighboring views share recognizable features; corners and connections have transition images.
Moving objects	People, cars and foliage take up little of the frame; enable people masking if needed.
File organization	Source data is backed up and readable, and the project folder has enough space.
Copyright / privacy	You have permission to capture and process the data; sensitive information is anonymized as agreed.

# 5 Automatic workflow: your first model

Best for first use and sample data. Drag a data folder or video onto the “Start 3D reconstruction” area on the right of the main window, or click the sample-data prompt to follow the sample guide.

Figure 3 The automatic workflow starts from the drop area on the right of the main window

Check your license and disk space first. On Basic, make sure the final training images will not exceed 600.

Drag in an image folder, a video, or a data folder with images and a sparse reconstruction; check the detected project on the wizard's data page.

For video, check the extraction FPS and panorama type. The wizard may pre-fill 2 FPS; on Basic with a video close to 10 minutes, change it to 1 FPS.

On the alignment page choose local alignment or suitable external results; consider people masking only if people appear.

On the training page choose the BSD ONE strategy, the maximum number of Gaussians, the number of iterations, and whether to use tiled training / sky masking. For the first run, keep the defaults and avoid changing several parameters at once.

Click “Start”. Watch the log and stage progress; when alignment and training finish, check the subject, edges and floaters in the 3D viewer before exporting.

Temporary projects: The automatic workflow may create a temporary project. When it finishes, choose a permanent location for the results and back up the project and model; do not use the temporary folder as a long-term archive.

# 6 Step-by-step workflow: projects and media

6.1 Create and load a project

Go to the “Data Import” page. Choose a location under “Project folder” and click “Create project”; to continue an old project click “Load project”.

Name projects with English letters, digits and hyphens, and keep them separate from system temporary folders; back up source data outside the project.

Make sure the main window log shows the project has loaded before importing data.

Figure 4 Data Import page: project folder, add files / folder, and log

6.2 Import photos, video and panoramic video

Click “Add files” to choose files, or “Add folder” to choose a data folder. You can also drag data onto the main window.

Check the import list and counts. When frame extraction settings appear for video, confirm the FPS; for native panoramic files such as .insv / .osv, confirm panorama mode.

Wait for extraction / undistortion to finish and check that the images are correctly oriented, without large black borders or heavy blur.

Check the final image count before alignment; on Basic it must not exceed 600.

Source folders: Some folder imports reference the original folder through a directory link. Do not move, rename or delete the source folder until the project is finished.

# 7 Camera alignment and external results

7.1 Local alignment

Go to “Alignment” → “Local alignment”.

If people in the footage interfere with alignment, tick “Remove people automatically”; it is usually unnecessary when no people appear.

Click “Start alignment”, watch the log and wait for the camera poses and sparse point cloud.

Check that the cameras form a continuous path and the subject is roughly complete. If many cameras are missing, first review image sharpness, overlap and extraction settings.

7.2 Import external alignment (Metashape / RealityScan / DJI Terra)

If you already have alignment results from photogrammetry software such as Agisoft Metashape, RealityScan or DJI Terra, you can import them into BSD Studio for 3DGS training and skip local feature matching and solving. Each tool exports a different file structure and coordinate system, so follow the steps below.

7.2.1 Export and import from Metashape

Align the photos in Metashape (you can prepare photos separately or use frames extracted by BSD, which are saved in the Data folder of the BSD project). When alignment is done, choose “File” → “Export” → “Export Cameras...”.

In the save dialog, set the file type to COLMAP text format (Colmap (*.txt)); name the file as you like, as shown below:

Figure 6 Choosing Colmap (*.txt) when exporting cameras from Metashape

In the export settings window, set the options as in the table below (see Figure 7):

Figure 7 Metashape camera export settings

Option	Recommended	Purpose
Export format	COLMAP (*.txt)	Exports standard camera intrinsics, extrinsics and the sparse point cloud
Save points	Yes	Writes sparse/0/points3D.txt, used to initialize Gaussian training
Export intrinsics	Yes	Writes cameras.txt with focal length and principal point
Export extrinsics	Yes	Writes images.txt with each view's quaternion and translation
Undistort	As needed (recommended if the original images have strong distortion)	Keeps pixel projection consistent with the pinhole camera model

Import into BSD Studio: the export folder contains two key folders, images and sparse. There are two quick ways to import:

Option 1 (drag and drop): drag the images and sparse folders together into a fresh BSD Studio main window. The app recognizes the standard COLMAP structure and opens the 3DGS training wizard.
Option 2 (from the UI): under “Alignment”, open “External alignment import”, choose “Reality Scan” in the source list (RealityScan and Metashape exports share the same structure), click “Browse”, select the folder and import.

7.2.2 Export and import from RealityScan

Complete alignment in RealityScan (or a RealityCapture workflow).

Then export in standard COLMAP format.

In the export settings, check these three key options (marked in red below):

Figure 8 Key COLMAP export settings in RealityScan

Key option	Value	Notes
Export format	COLMAP standard (.bin / .txt)	Exports a standard COLMAP sparse model
Export type	Binary (.bin) or Text (.txt)	BSD Studio reads both
Undistort images	Yes (True)	Undistorted images must be exported so camera poses and pixels line up exactly

Import into BSD Studio: the export folder also contains images and sparse folders. Select both and drag them into the main window, or use “Alignment” → “External alignment import” → “Reality Scan” and browse to the folder.

7.2.3 DJI Terra and LiDAR

DJI Terra: import the standard ZIP package exported from DJI Terra alignment, or the extracted folder. Camera distortion, POS attitude and the high-precision sparse point cloud are parsed automatically.

LiDAR scans: Pro can use colored LiDAR point clouds as the initial geometry, greatly improving Gaussian convergence in weakly textured or repetitive scenes.

License note: External alignment import (Metashape, RealityScan, DJI Terra, LiDAR, etc.) is a Pro feature. On Basic you will be prompted to upgrade. Keep the original image file names consistent with the external camera pose files, and do not edit the internal data structure by hand.

# 8 Gaussian training and evaluation

8.1 Parameters

Parameter	What it means	First-run advice
Densification strategy	Controls how Gaussians are added, split and pruned.	Start with BSD ONE as the baseline for comparison.
Max Gaussians	The upper limit on Gaussians during training; higher is not always better.	Start at 3 million and lower it if VRAM runs out.
Iterations	Number of optimization steps; more training does not necessarily remove structural artifacts.	Start with the default 30,000 and judge from the preview.
Tiled training	Splits a large scene to manage resources.	Enable only for genuinely large scenes or VRAM pressure.
Sky mask	Keeps the sky from being modeled incorrectly.	Enable for outdoor scenes with visible sky.

Figure 9 Gaussian training page: strategy, Gaussian count, iterations and progress

8.2 Start and monitor

Make sure alignment results are loaded and the training image count is within your license limit.

Set the parameters and click “Start Gaussian training”; watch the iteration progress, elapsed time and log.

To stop, use the app's pause / cancel controls; do not cut power or kill the process.

When training finishes, inspect from several angles: subject detail, back side, sky, floaters, large blobs near the camera and transparency.

Judging quality: More Gaussians or more iterations do not guarantee better quality. If floaters persist, investigate the data, alignment and strategy instead of only adding iterations.

# 9 3D viewer, editing and export

After training or loading a model, use the 3D viewer on the right to inspect it. The icon order follows the current version; the tools are described by function below so this guide does not depend on how the toolbar looks in a particular version.

Tool group	Purpose
Navigation & display	Reset view, rotation center, field of view, camera poses, grid, point cloud / Gaussian toggle
Measurement	Coordinates, distance, area
Animation	Add keyframes, preview the path, render video
Gaussian editing	Brush, lasso, invert, delete, undo
Output	Model export and mesh conversion

9.1 Cleaning floaters

Orbit around the subject in the viewer to tell real structure from background and floating Gaussians.

Select small areas with the brush or lasso, invert if needed, and confirm the selection before deleting.

After deleting, rotate to other angles to make sure the subject is intact; if you deleted too much, undo immediately.

Export a new result only when you are happy with the edits, and keep the original trained model as a fallback.

# 10 Saving, archiving and reproducibility

Treat a project as a traceable set of data: source data, preprocessing results, alignment, training parameters, training logs and final exports. Keeping only the final PLY is usually not enough to reproduce training or investigate problems.

Archive level	Keep
Source	Original photos / videos, camera and lens information, external alignment packages.
Project	Project folder, Data, AT, training configuration / checkpoints and related logs.
Delivery	Final Gaussians / meshes, thumbnails, version and usage notes.

Disk and paths: Stop training before archiving and make sure nothing is still writing in the background. After moving a project, first try “Load project” in the app and check that links to external source folders still work.

Recommended project naming

Use “Site/Object_CaptureDate_DataType_Version”, for example SubstationA_20260929_Photo_v01. Do not put personal contact details or confidential client information in shared export file names.

# 11 Buying Pro and managing your license

11.1 Before you buy

First check whether your project exceeds the Basic limit of 600 final training images, whether you need external alignment import, and whether your computer has enough VRAM.

Open “Buy Pro” on the official website and read the current prices, validity, terms and payment methods. Never pay through unknown third parties.

Choose monthly or yearly and complete checkout. Monthly adds 30 days and yearly adds 365 days, after any time you already have; the checkout page and terms are authoritative for prices and policies.

After payment, return to the app and wait for the license to sync, or reopen the license page, and check that it shows Pro and the expiry date. If it does not, keep your order number, payment time and account email, and contact support.

11.2 Purchase and refunds

Pro currently costs $19 per month (30 days) or $159 per year (365 days); taxes may be added at checkout depending on your country. We recommend evaluating your hardware and results with Basic and the free 14-day Pro trial first. Pro is activated on your account as soon as payment is confirmed. If you are not satisfied, you can request a full refund within 14 days of payment under our Refund Policy. Receipts are sent by email after payment.

Business purchasing and custom services: For multi-seat licenses, private deployment, offline licensing and custom development for research institutes, universities, survey and design institutes and mapping companies, email us for volume agreements and tailored solutions.

# 12 Support and troubleshooting

12.1 Contact

Website: https://www.bsd3dgs.com/en/ Email: bsd_technology@163.com Support hours: Monday–Friday 09:00–18:00 China Standard Time (UTC+8). Before reporting a problem, check the version number and system information under the app's Help menu so we can help you faster.

12.2 What to include in a report

App version, Windows version, GPU model / VRAM and GPU driver version.
Stage where the problem occurs: download, installation, sign-in code, import, frame extraction, alignment, training, 3D viewing or export.
Steps to reproduce, error screenshots, the time it happened and the related project log.
Data type, number of source files and number of final training images; for purchase issues, also the order number and payment time.
Anonymize sensitive client data first; do not send complete source data without permission.

12.3 Six things to check first

Symptom	Check first
Install / launch fails	Installer source, Windows permissions, security software, GPU driver.
No sign-in code received	Email address spelling, spam / junk folder, network, resend interval and daily limits.
Basic limit exceeded	Final training images in AT/images; video FPS and number of undistorted panoramic views.
Alignment fails / few cameras	Image sharpness, overlap, extraction interval, external alignment file structure.
Slow training or out of VRAM	VRAM usage, image / Gaussian counts, tiled training and other GPU tasks running in the background.
Floaters / blur	Check camera poses, masks and capture coverage first, then consider retraining or local editing.

# 13 FAQ

Q1 Why can a 10-minute video or 1-minute panoramic video still exceed the Basic limit?

Because the final training images are also limited to 600. A 10-minute video at 1 FPS gives about 600 frames; a 1-minute panoramic video at 1 FPS gives about 600 views after undistortion. A higher FPS or extra photos can exceed the limit.

Q2 I already have DJI Terra or RealityScan alignment. Do I need to align again?

No. Use it directly.

Q3 Floaters remain after more iterations. What should I do?

First check the capture angles, camera poses, moving areas and separation between background and subject. Floaters do not necessarily disappear with more iterations; clean small areas with the editing tools and keep the original model.

Q4 Can I continue a project on another computer?

Yes. Back up the project and any externally referenced source data first, then load it on the new computer to check. An account can be signed in on one computer at a time; sign out the old one on the Devices page, or contact support if you see a device limit.

Q5 Can older GPUs such as the GTX 10 series run BSD Studio?

Answer: yes. GTX 1060 (6 GB), GTX 1070, GTX 1080 and GTX 1080 Ti can all run local alignment and 3DGS training. Since the GTX 10 series usually has 6–11 GB VRAM, keep the maximum Gaussians between 1.5 and 3 million during training and update to the latest GPU driver.