Vimoo-1.0-Lite uses one ≈9.5B model for both text-to-video (T2V) and image-to-video (I2V).
- Diffusers Pipeline: Quick Start · Vimoo-1.0
- ComfyUI Workflows: ComfyUI-Vimoo · ComfyUI Weights · Lite ComfyUI Weights
- MLX (Apple Silicon): Vimoo-1.0-MLX · MLX Weights
🔥 News
- [2026-10] Release Vimoo 1.0: T2V / I2V models, Diffusers inference examples and ComfyUI workflows.
🎬 Showcase
🌐 Online API
Use Vimoo 1.0 directly via API. https://vimoo.ai
The model API includes three use cases: Vimoo 1.0, Vimoo 1.0 - Ecom, and Vimoo 1.0 - Lite.
🗺️ System Overview
The complete Vimoo system consists of the following three modules:
- Vimoo-PE: We developed a dedicated system that deeply interprets and refines incoming multimodal instructions, then transforms them into a format that Vimoo can easily process for generation. Vimoo-PE plays a crucial role in determining the quality of the final output; we highly recommend integrating it into your generation pipeline—or following the 'Prompt Guide' to build your own prompt system.
- Vimoo-Base: Generates 720p video based on the Vimoo-PE output.
- Vimoo-Enhancer: Feeds the 720p video to regenerate the output at 1080p resolution. It combines the resolution and clarity enhancement associated with generative repair of defects in appearance and motion, enabling it to produce high-resolution outputs with more accurate details and greater visual fidelity.
To help the community deploy Vimoo correctly, we provide three approaches.
- Open Platform API: Since Vimoo consists of three modules: Vimoo-PE, Vimoo-Base, and Vimoo-Enhancer, we offer the Open Platform API to support full end-to-end invocation of the complete pipeline, while the Enhancer is optional.
- OpenSource: We have open-sourced the Vimoo-Base model to enable validation of 720p output. For better results, you can refer to the 'Prompt Guide' section, which provides a detailed tutorial to help the community develop their own prompting systems. You may also further call the standalone Enhancer API to obtain high-definition 1080P results.
- Run Locally: We provide locally deployable models, including ComfyUI and MLX versions, to support running Vimoo-Base on lower-end devices. Here too, the 'Prompt Guide' section is essential.
In addition, we provide an e-commerce video generation API, specifically designed for product and advertising video generation. It follows the same three-stage pipeline and outputs a 20-second story-driven e-commerce video.
📦 Model Zoo
Vimoo 1.0 comes in two models, Vimoo 1.0 and Vimoo 1.0 - Lite. Each Hugging Face repository packages one of them for one runtime:
Each ComfyUI repository bundles its checkpoints, distilled LoRA adapters, shared encoders and VAE. Distilled models are available only for ComfyUI and MLX. Vimoo-1.0 T2V Distilled and I2V Distilled use different base checkpoints from Vimoo-1.0 T2V and I2V; download each task's matching base and adapter listed in the ComfyUI guide. The Vimoo-1.0 T2V Distilled, Vimoo-1.0 I2V Distilled and Vimoo-1.0-Lite TI2V Distilled adapters are not interchangeable. Parameters count the video denoiser only.
🚀 Quick Start
Installation
Use the Vimoo inference code on Linux with an NVIDIA CUDA GPU:
git clone https://github.com/Vimoo-AI/Vimoo-1.0.git
cd Vimoo-1.0
python -m pip install -r requirements.txt
python -m pip install "git+https://github.com/Vimoo-AI/diffusers.git"
See the installation guide for environment requirements.
The Vimoo-1.0-Lite TI2V example defaults to PyTorch SDPA, with no additional attention package. To select --attention-backend fa2, fa3, fa4 or sage, first install that backend's dependencies from the runtime options.
Download weights
hf download Vimoo-AI/Vimoo-1.0-Lite --local-dir ./models/Vimoo-1.0-Lite
On the first run, Diffusers downloads Qwen2.5-VL-7B-Instruct, the Wan 2.2 VAE and, for I2V, SigLIP from Hugging Face at the revisions recorded in modular_model_index.json. For offline use, run once with network access, then set HF_HUB_OFFLINE=1.
Generate video
Run from the inference repository root, replacing the image with your own file:
python examples/vimoo_1_0_lite_ti2v.py --model ./models/Vimoo-1.0-Lite --prompt 'A young woman in a red coat walks through a snowy city street and looks up at the falling snow.' --output lite-t2v.mp4
python examples/vimoo_1_0_lite_ti2v.py --model ./models/Vimoo-1.0-Lite --prompt 'The dog shakes its whole body, then sits back down in the same spot. The camera remains fixed throughout.' --image dog.png --output lite-i2v.mp4
Supplying --image selects image-to-video; omitting it selects text-to-video. The example outputs 1280 × 720, 121 frames at 24 fps. For Vimoo-1.0-Lite TI2V Distilled, use Vimoo-1.0-Lite-ComfyUI with the ComfyUI workflows, or Vimoo-1.0-Lite-MLX on Apple Silicon.
📝 Prompting Guide
Use English prompts. The current open-source models have been validated only with English prompts. Generation quality and prompt adherence in other languages have not been validated. We recommend translating prompts into English before generation.
Write clear, detailed prompts. Specific descriptions of the subject, scene, action sequence, and camera movement can help produce results that better match your intent than short, vague prompts. See the Prompt Guide for structure and examples, and focus on relevant, consistent visual details.
The Prompt Guide describes automatic prompt enhancement in the hosted service. Local examples use your input prompt directly; follow the guide's writing structure and examples when preparing prompts.
📖 Technical Architecture
Vimoo 1.0 uses a 40-block dual-stream MMDiT backbone. Joint attention connects the video and conditioning streams, combining prompt semantics, in-video text and optional reference-image features. A Wan 2.2 VAE encodes and decodes the video, while flow matching generates video latents.
Vimoo-1.0-Lite merges each 2 × 2 spatial region into one video token, reducing sequence length to one quarter while preserving the temporal axis, and restores the fine spatial grid before output projection. The denoiser has ≈9.5B parameters. See the Vimoo 1.0 README for the model architecture and inference components.
Highlights
- Prompt adherence: a dedicated text encoder turns prompts into precise generation conditions. An optional prompt enhancer can add visual, cinematic, temporal, and contextual detail, or be disabled to preserve the original wording.
- Cinematic visual quality: training on carefully curated aesthetic data, with detailed annotations for lighting, composition, contrast, color grading, and more, enables precise control over cinematic styles and a wide range of aesthetic preferences.
- Advanced motion generation: a larger and more diverse training corpus improves generalization across motion, semantics, and visual aesthetics, delivering leading performance among open-source and proprietary video generation models.
- E-commerce video generation: by faithfully highlighting product selling points and features while preserving fine-grained product details, it ensures strong product consistency across frames, delivering reliable and visually coherent results that meet the practical demands of e-commerce content creation.
- Efficient high-definition generation: the lightweight variant uses 32×32×4 spatiotemporal compression and supports T2V and I2V at 720p and 24 fps.
📈 Evaluation & Benchmarks
Generation Quality
Vimoo-VBench is a human evaluation developed by the Vimoo team, with 500 prompts covering advertising, social media, storytelling, education and entertainment. T2V is evaluated for instruction following, frame aesthetics, visual quality and motion; I2V also measures image-video consistency.
The tables compare model families on this evaluation set. Results for individual checkpoints depend on their weights and generation settings.
T2V human evaluation (%)
I2V human evaluation (%)
📁 Model Files
| Path | Purpose |
|---|---|
transformer/ | Shared T2V / I2V model |
scheduler/, guider/, guider_i2v/ | Sampling configuration |
recipe.json, inference_defaults.json | Recipe and defaults |
modular_model_index.json defines the required components.
Component provenance
The public model index records component locations and upstream revisions. Text conditioning uses Qwen2.5-VL and a Glyph encoder; image conditioning uses SigLIP, and video decoding uses the Wan VAE. Keep the bundled Glyph encoder and tokenizer with the model.
📜 License
Vimoo model weights in this repository are provided under the Apache License 2.0.
- Commercial use, modification, and redistribution are permitted under the terms of the license.
- Redistribution must include a copy of the license, retain applicable copyright and attribution notices, and identify modified files.
- Attribution and third-party component notices are provided in THIRD_PARTY_NOTICES.md. Third-party components remain subject to their own licenses.