Anything / Kl-f8-anime2 / Vae-ft-mse-840000-ema-pruned / Blessed / ClearVAE, but fp16/cleaned - smaller size, same result.
Short version of the description:
The size of the usual versions of these VAE is 320+ megabytes, VAE from here is 160, the result of generation on both versions is identical/practically identical (more details below). Each VAE is available in .safetensors and .pt extensions.
About each VAE:
kl-f8-anime2: my favorite, original
+ good for anime/art, high saturation and contrast, your images will definitely be bright
- in models with initially high saturation, saturation may become excessive, may distort some details just because of the high contrast
anything: my plan b, original
+ good for anime/art, moderate saturation, no overburning of the image, good for models with high saturation
- in models with initially low saturation, the picture may turn out to be faded
vae-ft-mse-840000-ema-pruned: for kl-f8-like, but different details, original
+ good for realistic/semirealistic, a slightly different balance of saturation and contrast relative to kl-f8-anime, the same advantages (but not the same result)
- same issues as in kl-f8-anime2
+ contrast version of anything vae
- more artifacts than if you use anything vae
+ makes sense if the others didn't fit or for your checkpoint is reccomended this VAE
- pale picture, can give more artifacts than the previous ones
As with all parameters - there is no universal and ideal, each specific one will suit better in a specific situation. Often you have to try each one, and compare.
The more boring part of the description:
The ideal way to save maximum space on your SSD or HDD: using these VAE, you can always use the "no vae" version of checkpoints, and manually select VAE in the web-ui.
a. saving space (checkpoints versions with VAE hang at least 2 gigabytes, "no vae" versions - 1.8~ = saving 160 megabytes from each model)
b. more control over the image style - the principle of VAE operation is largely similar to post-processing, so you can change "filters" in the form of VAE on the fly, and not be limited to one that is in the model itself (and most often it will be one of these VAE)
About the difference between regular VAE and fp16 VAE:
In many scenarios, there is no difference, but you can see minimal differences ("that house has a different window in the distance, a crease on clothes of a slightly different shadow tone", and so on) if you use certain hires.fix values. It happens when using hires.fix with "denoising strength" value above 0.3, I attached the last grid to demonstrate what differences between the usual VAE and fp16 VAE you can see (they're still not easy to spot right away) using high values - denoising strength: 0.5 / hires steps: 20.
About different VAE file names:
There are a huge number of VAE that are actually the same file that are constantly being renamed, so I chose 3 really different files (not considering Blessing2 / ClearVAE, because it's merges), and for me its enough of them for any image generation.
*my English is terrible, so feel free to point out mistakes in the text