This is the iteration of Luna VAE that adds contrast while maintaining the saturation of the previous version.
If there is a NaN error, the colors looks bad or any other thing, please let me know, thanks!
Description
Second iteration, seems working well.
Apparently in Pony, Noob and Illustrious models have a fairly high contrast, although it depends by the checkpoint..
FAQ
Comments (19)
Doesn't work with A1111. Solid black images, which seems to indicate a NaN error.
I checked A1111, ComfyUI, Forge, SDNext and on all of them it works, isn't your A1111 instance out of date or the downloaded file corrupted?
how do you train/ fine tune vae?
@karely_ai thank you
I tried it and it didn't change much with the realistic model, on the contrary it was a bit too much contrasty. I'm not sure if it would be better to train the model with that vae, but using it directly doesn't seem right.
Hi, regarding the contrast..
Yep, even I think it is too much at some times, honestly wouldn't recommend train with this vae bcuz of the fact that it could cause you to waste resources if the results are not the desired ones.
For inference, testing or mixing is ok, but training... I don't know.
@Karely_ai I think I will still try a little bit :)
@Karely_ai ok training with this vae is really not a good idea for 3d and realistic models
Increasing the contrast also reduces the spatial depth of the image, diminishes the sense of three-dimensional form, makes the image appear flatter, and causes the edges to feel more like 2D—therefore, it is not suitable for realistic or 3D images.
In InvokeAI (at least) this VAE produces an additive luminosity issue. Each image img2img pass gets progressively brighter. You actually lose contrast and get a washed out image.
ok a how good for image dragon and sex scene ?
I get these weird artifacts using this VAE as if there are tiny RGB lines all over the image.
Hi,
May I know which tool you use to generate images?
In case you are using a version lower than v0.3.0 of ComfyUI, please update your instance.
very nice
Sadly not FP16 safe and produces NaNs
what is diff if used a coutsom vae and if i didint?
good
good
i love this