Hugging Face Diffusers Adds Native Support for Nunchaku 4-bit Diffusion Inference

Hugging Face Diffusers now natively integrates Nunchaku Lite, a 4-bit quantization engine built on the SVDQuant method, allowing developers to load pre-quantized text-to-image models directly via a standard `from_pretrained()` call — no custom pipelines, separate inference libraries, or local CUDA compilation required. The integration also includes the `diffuse-compressor` toolkit for quantizing new model architectures and publishing them as regular Diffusers repositories.

Unlike weight-only quantization backends already supported in Diffusers (such as bitsandbytes, GGUF, and torchao), SVDQuant runs transformer layers with 4-bit weights and activations (W4A4), reducing both memory usage and inference latency. The approach handles the outlier problem common in diffusion transformers by offloading activation outliers into a small 16-bit low-rank weight branch. On an RTX 5090, a 1024×1024 image was reported to generate in approximately 1.7 seconds at around 12 GB peak VRAM, compared to roughly 24 GB for a standard BF16 pipeline.