INT8 Conv Rot, a novel quantization method, is gaining traction and is now natively supported in Comfy UI v0.27.0. This method offers significant performance benefits, particularly for GeForce RTX 20/30 series GPUs, and even surpasses FP8 formats on RTX 40/50 series.

INT8 Conv Rot addresses the limitations of traditional INT8 formats by employing a "Convolution and Rotation" technique to distribute outliers, thereby minimizing accuracy degradation. This innovation allows INT8 models to achieve performance levels previously only seen with FP8 formats, making it a strong candidate to become the standard for 8-bit quantized models.

Information on INT8 Conv Rot is still emerging, leading to some confusion. The method utilizes INT8 encoding with a specific "Conv Rot" quantization algorithm, differentiating it from other INT8 formats like INT8 Tensor-Wise. Support is actively being developed by Comfy-Org, with major models already being released in this format on platforms like Hugging Face.

Users can leverage INT8 Conv Rot through Comfy UI or Forge Neo. For optimal performance, it's recommended to ensure up-to-date Nvidia drivers and PyTorch installations. The native support in Comfy UI simplifies the integration of INT8 Conv Rot models, further solidifying its potential as the new industry standard.