Quality comparison: SVDQuant vs GGUF Q4 ?

[marvin-0042]

SVDQuant quality is great! the paper compares the quality with NF4, however GGUF Q4 (Q4_0, Q4_K_S) is another popular quantization in community and broadly used in ComfyUI with Image Gen model like FLUX.1 and Video Generation Model like Mochi 1. GGUF Q4 FLUX.1-schenell is also ~6.5GB with good quality, and it does not need big inference engine change.

Can we have any quality comparison between SVDQuant vs GGUF Q4 too, on FLUX.1 (and future Mochi 1)?

Today new image/video gen models like FLUX.1 or Mochi 1, community will first try ComfyUI + GGUF Q4 DiT to enable it on consumer GPU, especially low GPU RAM. If we can see SVDQuant is better than GGUF Q4, on FLUX.1 and/or Mochi 1, it will give community much bigger motivation to adopt or prioritize this very novel SVDQuant.