The proposed method is architecture agnostic, so, in principle, it works with any modern diffusion model. Given that the FLUX model is much larger, the benefits from its compression are more pronounced. However, the current implementation is focused towards SDXL-like architectures and adding support for FLUX would require significant effort. We hope to produce VQDM quantized FLUX models in the future.
Hi, @ninjasaid2k.
The proposed method is architecture agnostic, so, in principle, it works with any modern diffusion model. Given that the FLUX model is much larger, the benefits from its compression are more pronounced. However, the current implementation is focused towards SDXL-like architectures and adding support for FLUX would require significant effort. We hope to produce VQDM quantized FLUX models in the future.