Credits
Déjà View was built at SwampHacks XII, and all of our own code was written during the event. It stands on the open‑source models, libraries, tools, services and media below.
A licence is listed where our project records it. “See upstream” means we haven’t recorded it, and the project’s own licence applies.
Models
-
FlashWorldimlixinyang/FlashWorld, rev
424db448Photo mode: one photo, screenshot or sentence becomes a whole world.Apache-2.0 (code); weights: see upstream -
Wan2.2-TI2V-5BWan-AI/Wan2.2-TI2V-5B-DiffusersThe video model FlashWorld is built on (VAE, UMT5 text encoder, transformer config).Apache-2.0
-
Qwen2.5-VL-3B-InstructQwen/Qwen2.5-VL-3B-InstructFlashWorld’s captioner.See upstream
-
AnySplatInternRobotics/AnySplat, rev
5f5e208aMany photos become a splat world in one pass.MIT (code); weights: see upstream -
VGGTfacebook/VGGT-1BAnySplat’s geometry backbone.See upstream
-
Depth Anything V2Metric Indoor Small and Base, depth-anythingPer-frame depth for Video mode, the depth term when rebuilding a place, and lifting generated fill into 3D.See upstream
-
DeepLabV3torchvision, ResNet-50 and MobileNetPerson masks, so people don’t get baked into the room.See upstream
-
Big LaMaadvimman/lama, file from simple-lama-inpaintingInpainting the empty room behind people, and small fills.Apache-2.0
-
Qwen-Image-Edit-2511Qwen/Qwen-Image-Edit-2511Paints what the photos never saw (the ceiling, tree tops, the sky) and the back of each 3D person.Apache-2.0
-
Qwen2.5-VL-7BQwen-Image-Edit-2511’s text encoderDescribes what belongs in each gap before it is painted.See upstream
-
Multi-HMR 2NAVER, naver/multi-hmr2, commit
f066532dBody pose and shape per frame for the 3D people.Non-commercial licence (code and checkpoint), plus BEDLAM and AnnyOne terms -
Anny body modelNAVER,
anny0.6.0The body the 3D people are built on.Apache-2.0 code, CC0 assets -
DINOv3 backbone codeFetched by
torch.hubat run timeMulti-HMR 2’s image backbone.DINOv3 License -
SAM 2.1facebook/sam2.1-hiera-large and -smallObject masks.Apache-2.0
-
Grounding DINOIDEA-Research/grounding-dino-tinyText prompts to boxes for the object vocabulary.Apache-2.0
-
Optional image-to-3D props.MIT
No SMPL-family file is installed or used. Model weights are never committed to the project; each one is downloaded at setup.
Libraries and tools
In the web page and viewer
-
three.js
three0.186.1, with its GLTFLoader, meshopt decoder and RoomEnvironment add-ons3D rendering, and the headset on the landing page.MIT -
Spark
@sparkjsdev/spark2.3.1Gaussian splat rendering in the browser and the headset.MIT -
three-mesh-bvh0.9.15Fast collision and ray tests against the room.MIT
-
WebXR, Web Audio, MediaRecorderHeadset mode, spatial sound, recording.Web standards
Server
-
Bun and TypeScriptRuntime, server, bundler and tests.See upstream
-
@gradio/client2.7.1Fallback to the public FlashWorld Hugging Face Space.ISC
-
@google/genaiThe Gemini guide service.See upstream
Pipeline
-
Splat training, rendering and gap finding.Apache-2.0
-
spznianticlabs/spzThe compressed splat file format.MIT
-
COLMAP3.11Camera poses for photos and clips.See upstream
-
PyTorchRuns every model above.BSD-3-Clause
-
OpenCVImage and video processing.Apache-2.0
-
Open3D and pygltflibGeometry, texture baking and glTF export for the 3D people.MIT
-
torchvision; Hugging Face transformers, diffusers, accelerate, hub and safetensorsModel loading and inference.See upstream
-
NumPy, SciPy, Pillow, pillow-heif, plyfile, trimesh, warp-lang, einopsArrays, solvers, images (including HEIC), PLY files and meshes.See upstream
-
ffmpeg and uvVideo decode and encode; Python environments.See upstream
Roblox
-
Roblox Studio and LuauThe Roblox experience.Roblox terms
-
rojorojo-rbx/rojoRoblox build.MPL-2.0
-
lunelune-org/luneLuau test runtime.MPL-2.0
-
seleneLuau lint.See upstream
Testing
-
iwer2.5.0, Meta’s WebXR emulatorHeadset tests without a headset.MIT
-
playwright-core1.63.0Headless browser checks.Apache-2.0
-
pytest and bun testTest runners.See upstream
Services and hardware
- UF HiPerGatorRuns the models (Slurm, RTX PRO 6000 Blackwell GPUs).
- Hugging FaceHosts the model weights and the public FlashWorld Space we fall back to.
-
Google Gemini APIgemini-3.5-flash-lite, gemini-3.8-flash, gemini-3.8-flash-lite-ttsThe caption a Photo world is imagined from, the floating labels, spoken answers, narration and clip notes.
- Cloudflare TunnelThe public link.
- DeepSpaceThe memory board, live map and notes.
- RobloxThe multiplayer places.
- Auth0Sign-in in front of the routes that add, delete or drive things.
- Solana, Capital One Nessie, Tiger DataSponsor pieces, off unless configured: a devnet scoreboard, a sandbox tab and recorded replays.
- Meta Quest 3SThe headset we built for.
Media
-
“Quest 3S with controllers”3D model by OG_MATTERThe headset on the landing page. We kept only the headset and gave it new materials to match photos of the real one.
-
“Tilted Towers Tour | Fortnite OG S2”Free-camera footage by woodinjakeThe Tilted Towers world and the Brick House apartment world (71 frames).See upstream
-
Greyscale roomPublic demo splat from storage.googleapis.com/forge-dev-publicThe sample world in the gallery.See upstream
The Infinity Hall Game Lawn, the giant chess board and the kitchen are our own photos and clips.
Trademarks
Meta Quest, Roblox and Fortnite are trademarks of their owners; Déjà View is not affiliated with or endorsed by them.
AI tools used
- Google GeminiInside the product, as listed under Services and hardware.
- The generative models aboveMake up the parts of each world the camera never saw. The Reality check toggle tints everything they imagined.
Looked at, not used
- SuGaRAnttwo/SuGaRNo clear licence; 3DGS lineage is non-commercial.
- 2D Gaussian Splattinghbb1/2d-gaussian-splattingNo clear licence; 3DGS lineage is non-commercial.
- Gaussian Opacity Fieldsautonomousvision/gaussian-opacity-fieldsNo clear licence; 3DGS lineage is non-commercial.
- Gaussian Groupinglkeab/gaussian-groupingApache code, but the 3DGS dependency is non-commercial and the build is heavy.
- SegAnyGAussiansJumpat/SegAnyGAussiansApache code, same 3DGS dependency and CUDA build.
- FlashSplatflorinshen/FlashSplatNo clear licence.
- HUGSapple-aiml-research/ml-hugsNo clear licence; research use only.
- Gaussian splatting Blender add-onReshotAI/gaussian-splatting-blender-addonNo licence file.