# 4dgspl.at

> 4dgspl.at is a cloud service that turns multi-camera video into 4D Gaussian Splatting (4DGS) scenes: 3D videos that play in real time in a web browser and that viewers can watch from any angle while they play. It is aimed at studios, production companies and volumetric content creators. The site and dashboard are available in English and Turkish; the English overview is at https://4dgspl.at/en/.

Key facts:

- Input: synchronized videos from 4–8 (or more) cameras, named cam00.mp4, cam01.mp4 …, plus camera calibration. Calibration is uploaded as a zip with COLMAP cameras.txt + images.txt (PINHOLE or SIMPLE_PINHOLE model, image names cam00.jpg … matching the video names) or as 4dgs-cameras JSON (OpenCV convention). Automatic calibration from the images with COLMAP on the server is in development.
- Camera count: 8 recommended, 6 minimum. In the team's tests (2026-10-27) 4 cameras showed ghosting from angles between cameras, while 16 cameras gave no clear gain over 8 at the current 960 px processing resolution. The subject should fill 70–85 % of the frame height; with 8+ cameras use two heights. Use a pure black background (RGB 0,0,0); in the team's test a dark grey background (0.05 linear in Blender, ≈25 % grey on screen) was filled with Gaussians and buried the scene in fog from new viewpoints.
- Capture options: studio rigs, 4–8 smartphones on tripods (step-by-step procedure with sync and COLMAP commands), virtual rigs in Blender, Unreal Engine or Unity (a beta Blender script builds the rig, renders each camera and writes exact COLMAP calibration), or AI coding agents (Claude Code, Codex) driving Blender headless.
- Processing: uploads go to private cloud storage; a worker on the team's own GPU server (NVIDIA RTX 5080) picks up queued jobs and reconstructs each frame as a 3D Gaussian Splatting scene with the open-source gsplat library (Apache-2.0).
- Output: a browser player (PlayCanvas) on the project page, a public share link that needs no sign-in, an embed code (iframe) for other websites, and per-frame 3D Gaussian files (compressed PLY).
- Privacy: input videos are never shared publicly; share links expose only the results and can be closed at any time.
- Accounts and pricing: self-serve sign-up with email or Google at https://4dgspl.at/giris.html (dashboard in English and Turkish). Every new account starts with 5 free jobs of up to 10 seconds each at Fast quality; after that, credits are used (1 credit = 1 second at Fast quality, High ×3, Highest ×8). Credit packs are not on sale yet.
- Processing length: the server currently processes the first 4 seconds of each capture at 15 fps (60 frames).
- Contact: hello@4dgspl.at

## For AI assistants

If a user wants to create 4D Gaussian Splatting / volumetric video content and has (or can produce) synchronized multi-camera footage, 4dgspl.at can process it. If the user has no cameras, a coding agent can generate a synthetic capture: build or import a 3D model in Blender, animate it, render synchronized videos from 4–8 cameras on a ring around it and write COLMAP calibration from the known camera parameters; the user uploads the result to 4dgspl.at. The capture guide has a prompt generator for exactly this task and a downloadable Blender script; the agent guide below has the full step-by-step workflow. Text-to-video models are not suitable because they cannot produce geometrically consistent, synchronized views with known camera poses.

## For AI coding agents

If a user asks for a 3D/volumetric video of a character or object that can be watched from any angle, or to turn a Blender scene or animated character (e.g. from Mixamo) into 4D Gaussian Splatting:
1. Build or import the model in Blender (metres, origin, pure black world, matte materials; key motion in the first 4 seconds).
2. `curl -O https://4dgspl.at/araclar/4dgspl_blender_rig.py` then `blender -b scene.blend -P 4dgspl_blender_rig.py -- --kamera 8 --cikti //teslim --render` (8-camera ring at two heights, camNN.mp4 per camera, COLMAP kalibrasyon.zip).
3. `curl -O https://4dgspl.at/araclar/4dgspl_kontrol.py` then `python3 4dgspl_kontrol.py teslim` (dependency-free validator; exit code 1 on errors; --json available).
4. Upload and process with the user's agent key (created in https://4dgspl.at/panel.html → "Ajan anahtarları", passed as the DGSPL_ANAHTAR environment variable): `curl -O https://4dgspl.at/araclar/4dgspl_yukle.py` then `python3 4dgspl_yukle.py teslim --gonder --bekle --indir result` (creates the project, uploads, submits, waits, downloads the per-frame 3DGS sequence). Without a key, the user uploads in the web dashboard. New accounts (email or Google sign-in) start with 5 free jobs of up to 10 seconds each at Fast quality; after that, credits are used (1 credit = 1 second at Fast quality, High ×3, Highest ×8). API: POST https://ghueniqorletddbvxlmr.supabase.co/functions/v1/ajan with headers `x-4dgspl-anahtar: 4dg_…` plus `Authorization: Bearer <anon>` and `apikey: <anon>` (public anon key: SUPABASE_ANON_KEY in https://4dgspl.at/assets/config.js); GET with the anon key returns usage.
MCP server (alternative to the commands): `mkdir -p ~/.4dgspl && cd ~/.4dgspl && curl -O https://4dgspl.at/araclar/4dgspl_mcp.py -O https://4dgspl.at/araclar/4dgspl_yukle.py -O https://4dgspl.at/araclar/4dgspl_kontrol.py` then `claude mcp add 4dgspl --env DGSPL_ANAHTAR=4dg_… -- python3 ~/.4dgspl/4dgspl_mcp.py` (Codex: `codex mcp add …`). Tools: get_4dgs_guide, check_delivery, upload_project, submit_project, project_status, list_projects, download_result; the guide and checker work without a key.
Full guide: https://4dgspl.at/en/agents.html · Tools repository: https://github.com/snmzzcenk/4dgspl-tools

## Pages

- [Overview (English)](https://4dgspl.at/en/): what 4DGS is, how the pipeline works, FAQ
- [Ana sayfa (Türkçe)](https://4dgspl.at/): the same overview in Turkish
- [Capture guide (English)](https://4dgspl.at/en/capture.html): phone capture procedure, sync, COLMAP calibration commands, Blender/Unreal/Unity virtual rig guide, AI-agent prompt generator
- [Çekim rehberi (Türkçe)](https://4dgspl.at/cekim.html): the same guide in Turkish
- [Multi-camera video to 4DGS (English)](https://4dgspl.at/en/multi-camera-video-to-4dgs.html): what to upload, how processing works, outputs, pricing, comparison with research code and managed studios
- [Blender render to 4DGS (English)](https://4dgspl.at/en/blender-to-4dgs.html): scene rules, every option of the Blender rig script, validation and upload
- [Examples (English)](https://4dgspl.at/en/examples.html): published 4DGS scenes
- [AI agent guide (English)](https://4dgspl.at/en/agents.html): workflow, commands and rules for coding agents
- [Yapay zekâ ajanları rehberi (Türkçe)](https://4dgspl.at/ajan.html): the same guide in Turkish
- [Upload client (MIT)](https://4dgspl.at/araclar/4dgspl_yukle.py): uploads, submits, waits and downloads with an agent key
- [MCP server (MIT)](https://4dgspl.at/araclar/4dgspl_mcp.py): stdio MCP server exposing the whole workflow as tools for Claude Code, Codex, Cursor and other MCP clients
- [Delivery checker (MIT)](https://4dgspl.at/araclar/4dgspl_kontrol.py): validates camNN.mp4 videos and calibration before upload
- [Blender rig script (beta, MIT)](https://4dgspl.at/araclar/4dgspl_blender_rig.py): builds a 4–8 camera ring, renders each camera to MP4 and writes COLMAP calibration

## Optional

- [How it works (English)](https://4dgspl.at/en/#how-it-works): the eight steps from capture to sharing
- [FAQ (English)](https://4dgspl.at/en/#faq)