a collection of third-party finetuned vaes for: SD, SDXL, Flux.1, Flux.2, Qwen Image, Wan Video, others using the same latent spaces.
my criteria to quant/convert them:
- not trash: I will not convert vaes that are worse than the pretrained ones. I will not name them here not to adverstise.
- doesn't mean I wouldn't convert vaes that give less sharpness or saturation. I mean there has been a certain popular vae and it was the worst thing I've ever seen. don't ever fall for this, and actually test before taking it off from an experiment. my repo only has normal vaes, not "sharp" shit giving halos and noise everywhere every time.
usage:
everything for an initial generation, but for img2img/vid2vid, vaes that are not too diverged regarding their color preservation
*sd vaes (excluding the one for sd3.5, not here) go by the mit license
*distilled flux2 vae is here: https://huggingface.co/dummy9996/FLUX.2-small-decoder-bf16/tree/main
upd 7/27: flux2 hd was too saturated, so I've merged it with 2anime = flux2-animehd-mix-vae.safetensors
- Downloads last month
- 829
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for dummy9996/vaes-bf16
Merge model
this model