Practical overview
When it fits
Use Omnishow when a research or development workflow needs human-object interaction video generation with selectable text, image, audio and pose conditions, or needs the associated evaluation benchmark.
Who it is for
- Researchers studying human-object interaction video generation
- Developers adapting or reproducing the documented Wan-based generation workflow
- Teams evaluating multimodal human-object interaction video generation with HOIVG-Bench
Documented uses
- Generate human-object interaction videos conditioned on text and reference images.
- Add audio, pose video, or both to supported generation tasks.
- Train a Wan-based model using the provided training entry points and dataset metadata format.
- Evaluate human-object interaction video-generation methods with HOIVG-Bench.
Requirements & limits
The code is public on GitHub and is licensed under Apache-2.0. The documented workflow runs from the repository and requires downloading model assets and data.
- The documented inference workflow uses task-specific CSV fields, so input preparation depends on the chosen conditioning combination.
- The repository describes a Wan-based release and says OmniShow checkpoints are not included; inference with a fine-tuned checkpoint requires supplying one.
- The README identifies 480p and 720p settings and advises checking height and width for custom aspect ratios or resolutions.
- HOIVG-Bench is a separate evaluation resource, with 135 curated samples and task-specific metrics.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the Wan-based environment and assets
Set up the documented environment and download the model assets and example data needed for the later training or inference stages.
Before you start
- A clean Conda environment with Python 3.11
- Omnishow repository
- Wan backbones, tokenizer assets, Wav2Vec2 audio encoder and example dataset, downloaded using the repository scripts
- PyTorch installation matching the CUDA version in use
Steps
Clone the repository and create and activate a Conda environment with Python 3.11.
Install the project in editable mode and install its requirements.
If the default PyTorch build does not match the CUDA version, install a matching PyTorch build; the README gives CUDA 12.4 as an example.
Run the repository's weight and data download scripts.
Expected outcome
A prepared environment with the documented Wan-based assets and example dataset in the repository's expected directories.
Guide limitations
- This guide covers setup only; it does not establish that training or inference has been run.
- The release focuses on Wan-based reproduction and does not include OmniShow checkpoints.
Configure Wan-based inference inputs
Prepare the inference metadata and task settings that the documented Wan-based inference script reads.
Before you start
- The prepared Wan-based environment and downloaded model assets
- An inference CSV with fields matching the chosen generation task
- A fine-tuned checkpoint if using one
Steps
Choose a task: r2v, a2v, ra2v, rp2v or rap2v.
Populate the inference CSV with the fields for that task, such as text_prompt, reference-image paths, audio and/or pose_video.
Set the task and resolution in the inference script, and check height and width if using a custom aspect ratio or resolution.
If using a fine-tuned checkpoint, set DIT_CHECKPOINT in the inference script.
Expected outcome
Inference configuration ready for the documented script, which reads the metadata CSV and is specified to save generated videos under outputs/.
Guide limitations
- The repository does not include OmniShow checkpoints; using a fine-tuned checkpoint requires providing one.
- This guide describes documented configuration, not a verified generation run.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- Correr-Zhou/OmniShow · Official homepage title · retained in its source language