mochi-1-preview

Mochi-1 is a state-of-the-art open-source text-to-video generation model developed by Genmo. It achieves high-fidelity motion and strong prompt adherence in preliminary evaluations, significantly narrowing the gap between closed and open video generation systems.

Core components:

• Asymmetric Diffusion Transformer (AsymmDiT): A 10-billion-parameter diffusion model built on a novel architecture.
• Asymmetric Variational Autoencoder (AsymmVAE): Encoding-decoding model with an asymmetric structure
• Text Encoding: Uses a single T5-XXL language model for prompt encoding, avoiding reliance on multiple pretrained language models.

Key Features:

• Single-GPU Setup requires ~60GB VRAM for full operation.Recommended at least one NVIDIA H100 GPU for optimal performance. Supports multi-GPU operation with context parallel implementation. Memory Optimization options include CPU offloading, VAE tiling, and lower precision (bfloat16) variants to reduce VRAM usage.
• Current output resolution is 480p.
• The model supports generation of videos with photorealistic styles. Minor warping/distortions may occur in videos with extreme motion. Performs poorly with animated content.
• Organizations are advised to implement additional safety measures before deploying in commercial applications. 

---

The model is a component of the video generation pipeline, consisting of:

• Text encoder: ~4.8B parameters,
• Transformer: ~10B parameters,
• VAE: ~460M parameters.

Total: ~10.5B parameters

---

The 'VRAM requirements' entry is calculated based on the value provided by the authors (60 GB, fp32) with a multiplier corresponding to the assumed precision (e.g., 0.5 for 16-bit, 0.25 for 8-bit, and 0.125 for 4-bit).

Announce Date: 22.10.2024
Parameters: 0.01B
VRAM requirements: 7.5 GB using 4 bits quantization, 15.0 GB using 8 bits quantization, 30.0 GB using 16 bits quantization
Developer: genmo
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore mochi-1-preview capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.