For a separate human-centric audio-video project and its preparation format, use the daVinci-MagiHuman dialogue guide.
What Wan2GP and WanGP refer to
Wan2GP is a name you may encounter in tutorials, folder names and older links for the project now presented as WanGP. The current DeepBeepMeep repository describes a local interface for supported image, video and audio workflows. Start from that maintainer's current documentation before following an older installation video.
A runtime is different from a model. Selecting WanGP does not mean every job uses the same weights, inputs or sound capabilities. Choose the deliverable first: an edited image, a new shot, a spoken performance or a finished sequence. Then identify the supported model operation that can accept the required source material.
This guide supplies an original Night Window pilot and run-review materials. We read the public project documentation but did not install WanGP, download model weights, run its dry-run command or benchmark inference. Keep that distinction when comparing a documented workflow with your own working setup.
Choose a documented hardware route
| Decision | Current primary source | Record before a run |
|---|---|---|
| NVIDIA environment | Windows and Linux installation guide | GPU generation and matching runtime stack |
| AMD environment | Windows TheRock installation guide | Supported device and specified software route |
| Model fit | Models overview and selected operation | Weights precision memory profile and requested output |
| Batch handoff | Command-line queue documentation | Saved queue attachments and output directory |
The NVIDIA installation guide separates GPU generations and acceleration components. The AMD guide provides its own Windows TheRock route. Choose the documented path for your actual device rather than mixing installation commands from different generations or platforms.
A launcher filename does not establish support for a particular GPU backend. For a Mac image workflow, evaluate the separate LightX2V Mac-app route at its documented scope. Do not turn a platform name in a tutorial into a promise that every video model runs on that machine.
Interpret low-memory claims at the selected settings
The current models overview says select configurations can operate around 6 GB of VRAM and lists resolution, duration, references, precision and memory handling among the factors that affect fit. That is not a universal minimum for the entire catalogue.
Write down the device, model variant, requested size, length, input count and selected memory profile. If a job fails, retain the exact error and configuration before changing several settings at once. A smaller pilot can help establish whether the chosen route works, but a successful small run does not prove that a longer final sequence fits.
Keep preparation and loading time separate from generation and finishing. If you evaluate a local route against a hosted one, compare accepted deliverables and the work required to obtain them. A free code download is a different unit from the complete cost of an approved scene.
Prepare a Night Window pilot before a large batch
Our original pilot contains an amber cup with a single teal band, a coral wall phone and one rainy window. First prepare and approve a still reference. Then request one short, restrained shot using an input mode that actually supports that reference. Keep the object arrangement, materials and camera stable while allowing only the intended weather movement.
One continuous locked-off shot in an original quiet room at night. An amber cup with one teal band sits beside the rainy window. A coral wall phone stays fixed on the opposite wall. Preserve the cup shape, band, phone and room layout. Use restrained rain movement and a readable ending hold. No cut, zoom, new lettering or extra objects.
This is an untested creative brief, not a model-specific request payload. Select the real model, supported inputs and output settings in your installation. For a focus relationship between near and far details, use the three-plane composition plan rather than adding several camera effects to the first pilot.
Save the queue with the actual input material
The current command-line reference documents saved ZIP queues with attachments, settings JSON, an output-directory option and validation without generation. Create the queue in the working interface and preserve its real files. A hand-written review record is not a substitute for that runtime queue.
For an existing installation, the documented pattern below validates a saved queue before processing it. It is shown as a workflow reference and was not executed for this guide.
python wgp.py --process night-window.zip --dry-run
python wgp.py --process night-window.zip --output-dir ./night-window-outputsUse the environment and paths belonging to your actual setup. A successful validation is evidence about the queue's accepted structure, not about creative quality or whether every generation will finish. Inspect terminal results and the output directory after a real run.
Preserve a working setup when evaluating updates
Record the repository revision and environment used for an accepted pilot. When you evaluate an update, keep the same source and brief so you can identify a meaningful change in behavior. Retain the earlier working environment and its settings until the replacement has produced material you can inspect.
Follow the current maintainer instructions for updating the selected installation. Do not assume that a new code revision, a new model file and a changed acceleration backend are one indivisible change. Diagnose the failing layer before replacing everything. This article provides no universal repair command or tested migration result for your hardware.
Accept the exported media separately from job completion
Open the actual output file and check dimensions, duration, frame rate and sound tracks. Watch its complete action at normal speed. For Night Window, inspect the cup band, phone geometry, room arrangement, weather movement and final hold. Keep the first concrete defect and its time rather than approving a thumbnail.
For sound delivery, use the sound-bridge study. For captions, use the local SRT-to-VTT converter or the burned-subtitle workflow. Those finishing steps cannot establish that the generation itself preserved the scene.
Download the original pilot brief and blank run review sheet. The sheet leaves actual output, timing and acceptance empty. Develop original references and scene material through Clout AI Cinema or Image and Video as a separate creative route. This page does not establish a WanGP integration inside Clout.
