# OmniShow

> OmniShow is the only AI video generator purpose-built for human-object interaction video generation (HOIVG). It unifies text, reference image, audio, and pose conditions in one end-to-end model — producing studio-quality product demo videos with natural hand contact, frame-locked product consistency, and accurate lip-sync, all in a single generation pass. No studio, no crew, no filming required.

OmniShow is built on peer-reviewed research (arXiv: 2504.11804, published April 2026) by researchers from ByteDance, The Chinese University of Hong Kong, Monash University, and The University of Hong Kong. The model is fully open-sourced on GitHub and independently benchmarked on HOIVG-Bench — the field's first dedicated benchmark for HOIVG quality. It ranks #1 across all four generation modes against baselines including HunyuanCustom, HuMo-17B, VACE, Phantom-14B, and AnchorCrafter.

## Key Stats

- 4.9/5 rating from 1,200+ verified users
- 8,000+ active e-commerce sellers
- 2M+ videos generated
- Up to 10 seconds per clip in a single continuous pass

## Four Generation Modes

### R2V — Reference-to-Video
- **Inputs:** Text prompt + product photo + model reference image
- **Output:** Cinematic product demo video with stable hand-object contact
- Product color, texture, and shape are locked frame-to-frame — no drift, no distortion

### RA2V — Reference + Audio-to-Video
- **Inputs:** Text prompt + reference images + MP3 voiceover
- **Output:** Talking spokesperson video with frame-accurate lip-sync
- Covers pitch, pace, and natural pausing; no manual sync required

### RP2V — Reference + Pose-to-Video
- **Inputs:** Text prompt + reference images + pose sequence or video reference
- **Output:** Motion-controlled video matched to a defined pose path
- Supports full-body pose; no motion capture rig required

### RAP2V — Reference + Audio + Pose-to-Video (Industry First)
- **Inputs:** Text prompt + reference images + MP3 voiceover + pose sequence
- **Output:** Fully directed spokesperson video — appearance, audio, and motion locked from the first frame
- All four modalities processed together in one pass; no stitching, no consistency loss

## Additional Capabilities

- **Up to 10-second continuous clips** — no cuts, no stitching artifacts; long enough for a full product demo
- **Natural hand-object contact** — stable grip, natural finger wrap, realistic weight; no clipping or floating
- **Consistent character identity** — face, hair, outfit, and proportions stay identical across all frames
- **Talking avatar from one photo** — upload a portrait + audio track to generate a lip-synced avatar; no animation experience required
- **Cloud generation** — no GPU or software install required; 2–4 minute typical generation time
- **HD output at 720p, portrait-ready 9:16**

## How It Works (3 Steps)

1. **Upload reference images** — product photo and optional human model reference (JPG, PNG, WebP supported)
2. **Set generation conditions** — add any combination of text, audio (MP3), and/or pose sequence
3. **Generate and export** — receive a finished clip; preview, download, and publish directly

## Benchmark — HOIVG-Bench (April 2026)

HOIVG-Bench is the first benchmark designed specifically to measure HOIVG quality across four dimensions: visual fidelity, motion naturalness, identity consistency, and condition alignment.

| Model         | R2V       | RA2V      | RP2V      | Long-Shot  |
|---------------|-----------|-----------|-----------|------------|
| OmniShow      | ✓ Best    | ✓ Best    | ✓ Best    | ✓ Up to 10s|
| HunyuanCustom | Lower fidelity | Lower sync | —   | ✗          |
| HuMo-17B      | Lower fidelity | Lower sync | —   | ✗          |
| VACE          | Lower fidelity | —      | Lower adherence | ✗   |
| Phantom-14B   | Lower fidelity | —      | —         | ✗          |
| AnchorCrafter | —         | —         | Lower adherence | ✗   |

## Competitive Comparison

| Capability                          | OmniShow            | HeyGen        | Kling 3.0       | Runway Gen-4.5   | Seedance 2.0    |
|-------------------------------------|---------------------|---------------|-----------------|------------------|-----------------|
| Person holding & using product      | Purpose-built ✅    | Avatar only   | General motion  | Not addressed    | General motion  |
| All 4 inputs (text/image/audio/pose)| All four ✅         | 2 of 4        | 3 of 4 (no pose)| 3 of 4 (no pose) | 3 of 4 (no pose)|
| Stable hand & product contact       | Frame-locked ✅     | Avatar only   | Inconsistent    | Not addressed    | Not addressed   |
| Clip length                         | Up to 10s ✅        | Multi-minute  | Up to 15s       | 2–10s native     | Up to 15s       |
| Audio lip-sync                      | Full body ✅        | Full body     | 5 languages     | No native audio  | Native audio    |
| Pose / motion control               | Full body pose ✅   | None          | Ref video only  | Camera only      | None            |
| Product consistency across frames   | Locked ✅           | Varies        | Varies          | Varies           | Varies          |

## Target Users

- **E-commerce sellers** (Amazon, Shopify) — replace product video shoots; generate at catalog scale
- **TikTok Shop and social commerce brands** — 9:16 portrait-ready videos with automatic lip-sync
- **Short-form video creators and marketing teams** — full control over motion, product interaction, and character dialogue
- **AI researchers and developers** — fully open-sourced model weights and HOIVG-Bench for reproducible research

## Pages

- [Home](https://omnishowai.net/) — Overview, hero stats, and live gallery of generated videos
- [Gallery](https://omnishowai.net/#gallery) — 20+ AI-generated demo clips across all four modes (R2V, RA2V, RP2V, RAP2V)
- [Features](https://omnishowai.net/#features) — Detailed breakdown of all four generation modes with input/output specs
- [Benchmark](https://omnishowai.net/#benchmark) — HOIVG-Bench results and competitive comparison tables
- [How It Works](https://omnishowai.net/#how-it-works) — Three-step workflow explanation
- [FAQ](https://omnishowai.net/#faq) — 9 frequently asked questions covering inputs, pricing, comparisons, and research

## Research & Open Source

- [arXiv Paper](https://arxiv.org/pdf/2604.11804) — Full technical paper (April 2026)
- [GitHub Repository](https://github.com/Correr-Zhou/OmniShow) — Open-source model weights and code
- [HOIVG-Bench Dataset](https://huggingface.co/datasets/donghao-zhou/HOIVG-Bench) — Benchmark dataset on Hugging Face

## Contact

- Support: support@omnishowai.net
- Website: https://omnishowai.net