Inside WanGP
WanGP is not just a launcher. Its source combines memory-aware model loading, quantization, queues, media preparation, finishing tools and local automation into one coherent workspace.
01 / MORE HARDWARE
WanGP’s implementation adds the execution layer that turns a collection of open models into a practical local system.
WanGP separates video, image and audio memory profiles and adapts offloading, pinned memory and asynchronous transfers to the machine. The goal is to keep the GPU busy without pretending every model fits every card.
Select models start at 6 GB VRAMThe runtime routes multiple checkpoint formats through dedicated handlers, including scaled FP8, NV FP4, bitsandbytes NF4, Nunchaku INT4/FP4, INT8 ConvRot and GGUF.
int8 · fp8 · GGUF · NF4 · NV FP4Model definitions and download registries let WanGP select and fetch compatible transformer, encoder, upsampler and processor files rather than forcing one universal package on every GPU.
Models arrive on demandThe project documents paths for older NVIDIA generations and current RTX hardware, plus dedicated AMD setup for supported RDNA 2, 3, 3.5 and 4 families.
NVIDIA + AMDThe queue is a real workflow surface: jobs can be edited, paused, autosaved to portable bundles and processed without the web interface through the command line.
Save · resume · batchWanGP exposes local API, MCP and plugin surfaces for private automation. Plugin model extensions, upsamplers and audio processors are discovered and registered at startup.
Local automation surfacesUnder the hood
The main application registers dedicated quantization handlers, maintains separate media memory profiles, autosaves queue bundles, discovers plugin model extensions and registers spatial, temporal and audio processors.
02 / CREATION TOOLS
Every stage—from preparing an input to finishing the soundtrack— stays within reach of the same local workspace.
A connected workflow
Inputs and outputs do not disappear between tools. Galleries, metadata, settings bundles, selected media and queue attachments make each result a starting point for the next operation.
Draw or expand masks, remove backgrounds, crop and resize references, extract pose, depth or optical flow, separate speakers, and remove noise or songs before generation.
Per-model prompt enhancers, reusable settings, macros, reference images and inline sliding-window commands help turn a concept into repeatable generation instructions.
Advanced video, image and audio galleries retain creation context. Browse results, inspect metadata, restore settings and feed an output directly into the next operation.
Temporal and spatial processor registries power frame interpolation and upscaling. Audio processors can build soundtracks, replace voices or remux existing video with a new track.
The low-VRAM offline assistant can orchestrate generation and media utilities, including transcription, video splitting and color-frame creation, from the interface or CLI.
The plugin catalog spans galleries, finetune and LoRA management, image workbenches, prompt libraries, queue notifications, timeline tools and additional model or processor types.
Audio is first-class
Generate music or speech, inspect audio in its own gallery, add sound to a video, replace voices and remux tracks without leaving WanGP.
Processing documentation
Ready to create
