Wan 2.1 stands out in tasks such as text-to-video (T2V), image-to-video (I2V), and video editing, while also supporting multilingual visual text generation. It is optimized for consumer-grade GPUs, with the T2V-1.3B model requiring only 8.19 GB of VRAM.
Mochi 1, an open-source AI model, excels in high-fidelity video generation with impressive motion quality and strong prompt adherence. Although it can run on a single GPU, it demands approximately 60 GB of VRAM for optimal performance.
Video generation models are rapidly evolving, providing users with the ability to create high-quality videos from text prompts or images. These models vary in architecture, capabilities, and hardware requirements, making it essential to understand their strengths and limitations. Two prominent models in this space are Wan 2.1 and Mochi 1.
Start a free trial on Novita AI today. To integrate the Wan 2.1 API, visit our developer docs for more details. Moreover, we provide the full-powered 14B version.
Novita offers highly competitive pricing in the market.
For example, a Wan 2.1 720P 5-second video costs only $0.4 per video
while a similar video on Replicate costs $2.39 per video
We’re now testing the two models by inputting the same text prompts to evaluate their understanding of the text and the final output of the videos.
Prompt: A garden comes to life as a kaleidoscope of butterflies flutters amidst the blossoms, their delicate wings casting shadows on the petals below. In the background, a grand fountain cascades water with a gentle splendor, its rhythmic sound providing a soothing backdrop. Beneath the cool shade of a mature tree, a solitary wooden chair invites solitude and reflection, its smooth surface worn by the touch of countless visitors seeking a moment of tranquility in nature’s embrace.
Wan 2.1
Mochi
Prompt: A golden retriever, sporting sleek black sunglasses, with its lengthy fur flowing in the breeze, sprints playfully across a rooftop terrace, recently refreshed by a light rain. The scene unfolds from a distance, the dog’s energetic bounds growing larger as it approaches the camera, its tail wagging with unrestrained joy, while droplets of water glisten on the concrete behind it. The overcast sky provides a dramatic backdrop, emphasizing the vibrant golden coat of the canine as it dashes towards the viewer.
Generates videos at 480P resolution, with 720P support planned for future updates.
Capabilities
Excels in Text-to-Video (T2V) and Image-to-Video (I2V) tasks.
Primarily a Text-to-Video (T2V) model; I2V implementation has been requested by the community.
Video Length
Generate a 5-second 480P video on an RTX 4090 in about 4 minutes.
Generates videos up to 5.4 seconds in duration. Actual testing may take less than 1 minute to generate it.
Architecture
Wan 2.1
Wan 2.1 is built on a diffusion transformer paradigm, enhanced by the Flow Matching framework.
It employs Wan-VAE, a cutting-edge 3D variational autoencoder that ensures efficient compression and high fidelity in motion reproduction.
A T5 encoder enables the processing of multilingual textual input seamlessly.
The architecture integrates an advanced parameter modulation system to optimize the prediction and incorporation of textual information into generated videos.
Cross-attention mechanisms within each transformer block embed textual input directly into the model’s structure, enhancing alignment and context integration.
Mochi 1
Mochi 1 is powered by a 10-billion-parameter diffusion model built on the Asymmetric Diffusion Transformer (AsymmDiT) architecture.
It features an asymmetric encoder-decoder structure, enabling highly efficient and high-quality compression.
The AsymmVAE compresses videos by a factor of 128, achieving 8×8 spatial and 6x temporal compression into a 12-channel latent space.
A single T5-XXL language model is used to encode prompts, ensuring robust language understanding and integration.
The architecture is designed to streamline text processing, allowing the model to allocate more neural capacity to visual reasoning and video generation.
Hardware Requirements
Wan 2.1
The T2V-1.3B model requires only 8.19 GB VRAM, making it compatible with consumer-grade GPUs.
For example, the generation of a 5-second 480P video takes about 4 minutes on an RTX 4090.
Mochi 1
Requires ~60GB VRAM for single GPU operation.
It supports both multi-GPU operation and single-GPU operation.
Initial reports suggested needing 4 H100 GPUs, but optimizations have reduced this significantly like 1 GPU.
Application
Wan2.1
Suitable for diverse businesses utilizing AI for developing high-quality visual content in a cost-effective manner.
Applicable in creative and professional contexts due to its ability to produce text content directly within videos.
Mochi 1
Designed to help creators quickly turn written content into video, without needing extensive editing skills or equipment.
Versatile applications in research, product development, and creative expression.
Conclusion
Choose Wan 2.1 if you need a versatile model that supports multiple tasks (Text-to-Video, Image-to-Video, video editing), multilingual capabilities, and efficient performance on consumer-grade GPUs. It is especially well-suited for applications requiring high performance in dynamic motion, spatial relationships, color accuracy, and multi-object interactions.
Opt for Mochi 1 if your focus is on high-fidelity motion and strong prompt adherence in video generation. While it has higher VRAM requirements, its open-source nature and compatibility with tools like ComfyUI make it an excellent choice for creative experimentation and research.
Novita AI is the All-in-one cloud platform that empowers your AI ambitions. Integrated APIs, serverless, GPU Instance — the cost-effective tools you need. Eliminate infrastructure, start free, and make your AI vision a reality.
