dc-ai-projects/DC-Gen

What it solves

DC-Gen addresses the high computational cost and low throughput of modern visual diffusion models, which often makes high-resolution (e.g., 4K) image and video generation prohibitively expensive in terms of training and inference time.

How it works

DC-Gen is an acceleration framework that transfers pre-trained diffusion models into a deeply compressed latent space. It uses a technique called Embedding Alignment to transfer the base model's knowledge to the new latent space, allowing the model to generate visuals with correct semantics without needing to retrain the diffusion model's weights from scratch. The process is completed with a few steps of end-to-end fine-tuning or LoRA fine-tuning to recover full quality.

Who it’s for

Developers and researchers working with diffusion models who need to achieve significant speedups in inference and the ability to generate high-resolution images and videos without the massive overhead of original base models.

Highlights

• Massive Speedup: Achieves up to 53.8× faster inference on H100 GPUs at 4K resolution for models like FLUX.1-Krea.
• Versatile Application: Supports text-to-image (1K and 4K), text-to-video, image-to-video, and instruction-based image editing.
• Efficient Adaptation: Enables rapid adaptation of autoencoders without requiring full model retraining.
• High Quality: Maintains the realism and text-rendering capabilities of the base models while drastically reducing token redundancy.