designfuchen/worldforge
Media & DesignA macOS studio app that drives Colab or local GPUs as a remote backend for 3D world and AIGC generation.
What it does
WorldForge is a hybrid pipeline that pairs a native macOS SwiftUI studio app and a local Python daemon with a remote GPU backend on Google Colab G4 (RTX PRO 6000 Blackwell, 98 GB VRAM) or a local RTX 4090/5090 machine. A single multiplexed SSH tunnel connects the Mac to a FastAPI orchestrator on the GPU host, which keeps model workers resident so image and video generations do not reload weights each run.
The project targets creators who generate 2K stills and 3D fly-through video on rented or local NVIDIA hardware but want a responsive desktop client. Built-in presets cover text-to-image, text-to-video, image-to-video, TTS and NVIDIA Lyra long-flythrough modes; arbitrary Hugging Face repos can also be registered. On the Mac side, Apple Silicon post-processing handles frame interpolation, speed changes, scaling and end-frame holds.
Quick start
Prerequisites: a Mac (Apple Silicon recommended), ffmpeg installed via Homebrew, and the colab CLI logged in with available compute units. For the local GPU option, no Colab CLI is needed.
git clone https://github.com/designfuchen/worldforge.git
cd worldforge
open WorldForge.app # starts the native client and the background forge.py daemon
Alternatively run python3 forge.py --open to launch only the local daemon and Web workbench. For a local NVIDIA host, start the backend there with ./start_local_gpu.sh (equivalent to python3 remote/forge_server.py --local --host 0.0.0.0 --port 17860), then connect from the Mac with python3 forge.py --gpu-url http://<host-ip>:17860 --open. The generated assets land in ~/Movies/WorldForge.
Getting a usable Lyra fly-through
The README devotes a section to reproducing long "always-moving" shots: bidirectional generation with reversed-segment stitching, a frozen-tableau prompt template, DA3/MoGe inverse-depth alignment, center-cropping to the native 832:480 aspect ratio, and RIFE interpolation on the Mac. These are project-specific presets and methods, not generic defaults, so expect to follow the documented mode and parameter values in the Lyra panel rather than a bare prompt.
Constraints
A Google Colab account with compute units is required for cloud mode; billing and idle shutdown are managed through the app's watchdog and status endpoints. The native macOS client is macOS-only, while the Web UI and daemon are Python 3 standard-library based. Local RTX 4090/5090 mode needs a Linux or WSL2 machine reachable over the network. Models are large (tens of GB), so RAM, VRAM and disk on the GPU host matter; the project documents automatic VAE tiling and offloading on smaller cards. License is MIT.
HotRepo's take
The repository is young and entered HotRepo's new-project list on 2026-11-09 with 101 stars and 33 forks in its first growth window. There is no long-term star curve yet, so no trend claim is warranted. What stands out is the unusual split between a native SwiftUI client, a zero-dependency local daemon and a remote worker pool with pinned-memory model swapping, which is a more specific engineering claim than a typical wrapper script. Whether that complexity pays off for a given creator depends on their access to a G4-class Colab runtime or a local 4090/5090 host.
Written by AI from the project's README, updated . The project's own docs are the source of truth.