Researchers at Tarbiat Modares University have introduced DuoDiT, a dual-stream architecture for fine-tuning diffusion transformers. The method achieves competitive image generation results on ImageNet-1K while updating only 2.6% of the model's parameters.
This technical development addresses the high computational cost of adapting large-scale diffusion transformers for specific visual tasks. By freezing the primary backbone and utilizing a secondary stream for local refinement, developers can achieve high-fidelity image generation without the massive hardware requirements typically associated with full model retraining. For the streaming industry, this efficiency is critical as generative AI production costs move toward high-resolution video synthesis and personalized content creation. Reducing the trainable parameter count to 2.6% allows smaller engineering teams to fine-tune state-of-the-art models on consumer-grade hardware. Watch for whether this dual-stream approach is integrated into open-source video generation frameworks like Stable Video Diffusion to accelerate custom model deployment.
The shift toward parameter-efficient fine-tuning is part of a broader industry trend where AI video creation workflows are increasingly prioritizing narrative continuity and resource optimization. As these integrated newsroom AI workflows become standard, the ability to fine-tune models with minimal overhead will be a key competitive advantage. This is especially relevant as enterprise AI video generation continues to scale content output to bypass production bottlenecks, a trend further supported by AI video generation platforms entering professional pipelines.
Researchers at Tarbiat Modares University have developed DuoDiT, a dual-stream architecture for diffusion transformers that updates only 2.6% of model parameters. By freezing the primary backbone and adding a trainable secondary stream, this method achieves high-fidelity image generation while significantly reducing the computational costs and hardware requirements for fine-tuning large-scale models.
DuoDiT is a dual-stream fine-tuning method for diffusion transformers that keeps the primary backbone frozen while using a trainable second stream to capture fine-grained texture and boundary details.
DuoDiT updates only 17.95 million parameters, which accounts for 2.6% of a 690.39 million parameter model.
DuoDiT achieved a competitive FID of 2.219 on ImageNet-1K and recorded the lowest KID and LPIPS-Mean scores among the evaluated parameter-efficient fine-tuning techniques.
It allows smaller engineering teams to fine-tune state-of-the-art models on consumer-grade hardware, which is critical for the future of high-resolution video synthesis and personalized content creation.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source