Why do input types matter for Gaussian Splatting?
Input quality determines how much visual information the model has to reconstruct the subject.
Gaussian Splatting depends on repeated visual observations. If the input does not show an area clearly, the final model may not reconstruct that area well. This is why capture quality matters as much as software choice.
Different inputs solve different problems. Photos can be sharp and controlled. Video can be fast and convenient. 360 footage can capture spatial context. Datasets can support repeatable production workflows.
When should you use images for Gaussian Splatting?
Use images when you want maximum control over sharpness, lighting, and coverage.
Image input guide
| Best use | Why images help | Capture advice |
|---|---|---|
| Products | Sharp photos can show labels, edges, and material detail. | Shoot full orbit and detail passes. |
| Small objects | Camera position can be controlled carefully. | Avoid extreme closeups without overview shots. |
| Art and collectibles | Still images preserve detail. | Use stable lighting and avoid glare. |
| Controlled studio captures | Lighting and background can be consistent. | Keep exposure and distance stable. |
| First tests | Easy to inspect and remove weak frames. | Use more angles than you think you need. |
When should you use video for Gaussian Splatting?
Use video when speed and continuous coverage matter, especially for objects, rooms, and walkthroughs.
Video input guide
| Best use | Why video helps | Capture advice |
|---|---|---|
| Quick product capture | A slow orbit can cover many viewpoints quickly. | Move slowly around the product. |
| Rooms | Video captures continuous spatial movement. | Walk slowly and avoid sudden turns. |
| Outdoor scenes | Video can capture broad context. | Avoid motion blur and rapid exposure changes. |
| Client tests | Fast capture helps early proof-of-concept work. | Review quality before final production. |
| Social or creator workflows | Convenient capture from common devices. | Prioritize stability over speed. |
When should you use 360 video or 360 captures?
Use 360 capture when the subject is a space, interior, property, or environment where surrounding context matters.
360 capture can be useful because it records more spatial context than a narrow field-of-view video. This is especially relevant for real estate, interiors, tourism, showrooms, construction documentation, and large scenes.
The same quality rules still apply. Move smoothly, keep exposure stable, and make sure important transitions are captured clearly.
360 input guide
| Use case | Why 360 helps | Risk to watch |
|---|---|---|
| Real estate | Captures rooms and surrounding context. | Missing transitions between rooms. |
| Showrooms | Shows product placement and environment. | Fast camera movement. |
| Construction documentation | Captures spatial state quickly. | Low texture or poor lighting. |
| Tourism and locations | Communicates place and atmosphere. | Large file size and performance. |
| Large interiors | Reduces need for many narrow passes. | Exposure changes near windows. |
When should you use a prepared dataset?
Use a prepared dataset when you need a more organized, repeatable, or advanced workflow.
Datasets are useful for teams that already have structured image folders, extracted frames, or repeat capture pipelines. They can make production more consistent, especially when many models are created in the same way.
The advantage of a dataset is organization. The risk is assuming that organization alone guarantees quality. Bad inputs in a neat folder are still bad inputs.
Dataset workflow guide
| Dataset need | Why it matters | Best practice |
|---|---|---|
| File organization | Teams need to find and review inputs. | Use clear names and folders. |
| Input filtering | Bad frames can reduce quality. | Remove blur, duplicates, and irrelevant media. |
| Repeatability | Production workflows need consistency. | Use the same capture rules across projects. |
| Scale | Many products or spaces require structure. | Track client, subject, date, and version. |
| Review | Datasets can hide weak files. | Inspect before training. |
What makes an input good or bad?
Good inputs are sharp, overlapping, complete, consistent, and relevant to the final model.
Input quality checklist
| Quality factor | Good input | Bad input |
|---|---|---|
| Sharpness | Details are readable. | Motion blur or heavy compression. |
| Coverage | All important areas appear from multiple angles. | Missing sides, corners, backs, or transitions. |
| Lighting | Stable and consistent. | Flicker, harsh shadows, strong exposure jumps. |
| Relevance | Media focuses on the subject. | Unrelated frames or distracting movement. |
| Overlap | Neighboring views connect naturally. | Large viewpoint jumps with little shared detail. |
| Subject stability | Object or space stays consistent. | Moving objects or changing layout during capture. |
Frequently asked questions
Can Splatware use photos for Gaussian Splatting?
Yes. Photos are often a strong input when they are sharp, consistent, and cover the subject from many overlapping viewpoints.
Can Splatware use video for Gaussian Splatting?
Yes. Video can work well when it is slow, stable, and complete. Fast or blurry video usually produces weaker results.
Can Splatware use 360 video for Gaussian Splatting?
Yes. 360 video can be useful for spaces, interiors, real estate, showrooms, and larger environments where surrounding context matters.
Can Splatware use INSV files?
INSV-style workflows can be useful when working with Insta360-style 360 capture, but footage may need to be prepared or converted correctly before training.
What is the best input for product splats?
Controlled photos or slow product video are usually best. Products need sharp detail, consistent lighting, and complete coverage from all important angles.
What is the best input for real estate splats?
Slow walkthrough video or 360 capture is usually best. The capture should cover every room and transition with consistent lighting.
Should I upload every file I captured?
Not always. Remove clearly bad inputs such as blurry frames, accidental clips, irrelevant images, or files that do not help reconstruct the subject.
What is the biggest input mistake?
The biggest mistake is incomplete coverage. Even high-quality footage can produce weak results if it does not show every important area from enough overlapping angles.