Idea
API to accelerate diffusion transformer models for faster image and video generation benefiting AI developers and content creators
Research Paper
Core Innovation
This paper introduces Cluster-Driven Feature Caching (ClusCa), which clusters spatial tokens each timestep and computes only one token per cluster, propagating information to others. This reduces token computation by over 90% without retraining the diffusion transformer. ClusCa achieves up to 4.96x speedup with minimal quality loss, applicable to any diffusion transformer model.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing demand for AI-generated media and efficient model deployment in creative industries and enterprises.
Potential Customers & Pain Points
- AI Developers Needing Faster Model Inference
- Content Creators Requiring Efficient High-Quality Image and Video Generation
- Enterprises Deploying Diffusion Models with Limited Compute Resources
Business Model
Offer ClusCa as a SaaS API and SDK for integration into existing AI pipelines with tiered pricing based on usage and enterprise features
Competitive Landscape
- Runway ML
- Stability AI
- OpenAI
Implementation Challenges
- Integration with diverse diffusion transformer architectures
- Maintaining quality with aggressive token reduction
- Adoption by established AI model providers
Validation Strategy
- Benchmark speed and quality on popular diffusion models
- Pilot integration with AI content creation platforms
- Collect user feedback on performance and quality trade-offs
Research Paper Overview
Compute Only 16 Tokens in One Timestep: Accelerating Diffusion Transformers with Cluster-Driven Feature Caching
Summary
Diffusion transformers generate high-quality images and videos but are computationally expensive due to iterative denoising. Existing feature caching accelerates by reusing features across timesteps but ignores spatial similarity. This paper introduces Cluster-Driven Feature Caching (ClusCa), which clusters spatial tokens each timestep, computes one token per cluster, and propagates information to others, reducing tokens by over 90%. ClusCa works on any diffusion transformer without retraining and achieves up to 4.96x speedup with minimal quality loss, validated on DiT, FLUX, and HunyuanVideo for text-to-image and text-to-video generation.