"As good as SDXL VAE but runs twice as fast and uses significantly less memory." https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/discussions/7
"Same license on stable-diffusion-xl-base-1.0
same vae license on sdxl-vae-fp16-fix
Troubleshoot:
Do not use the refiner with VAE built in
Try launch param: --medvram --opt-split-attention --xformers
SDXL-VAE-FP16-Fix is the [SDXL VAE](https://huggingface.co/stabilityai/sdxl-vae, but modified to run in fp16 precision without generating NaNs.
Details:
SDXL-VAE generates NaNs in fp16 because the internal activation values are too big:
SDXL-VAE-FP16-Fix was created by finetuning the SDXL-VAE to:
keep the final output the same, but
make the internal activation values smaller, by
scaling down weights and biases within the network
There are slight discrepancies between the output of SDXL-VAE-FP16-Fix and SDXL-VAE, but the decoded images should be close enough for most purposes." - bdsqlsz
NOT MY WORK - REUPLOADED HERE FOR EASE OF USE
COMMISSIONS NOW ACCEPTED!
I have been away saving to upgrade my pc, please help fund more work:
Description
FAQ
Comments (28)
I get errors now
in that case, try apply settings or use the default one, sorry!
- Do not use the refiner with VAE built in
Try launch param: --medvram --opt-split-attention --xformers
Error
in that case, try apply settings or use the default one, sorry!
- Do not use refiner with VAE built in
My Launch param: --medvram --opt-split-attention --xformers
working fine
错误
错误
错误
Edit: the one from HF works
I have lots of errors trying to load this one.
RuntimeError: Error(s) in loading state_dict for AutoencoderKL: Missing key(s) in state_dict: "encoder.down.0.block.0.norm1.weight", "encoder.down.0.block.0.norm1.bias", "encoder.down.0.block.0.conv1.weight", "encoder.down.0.block.0.conv1.bias", "encoder.down.0.block.0.norm2.weight", "encoder.down.0.block.0.norm2.bias" [and a lot more]
try change yaml to in_channels: 9
it;'s the SAME file!
Good luck
I have not been able to get this to work, just get "TypeError: 'NoneType' object is not iterable"
Is this something I am doing wrong?
Is it possible to add the file as a download directly on this Civitai page? I'd like to be able to pull this in automatically with Stability Matrix, but have to do it manually because the download is only on huggingface
It was there but somehow dissapeared, it's back uploaded here now!
If this applies fixes and improved performance without any downsides, why did Stability AI decide to ship an FP32 VAE by default?
Cuz fp16 is less precise and just a bit worse. Thats all. In case of low-end configuration small inaccuracies are the sacrifice that worth the 2x speed and free VRAM
If code was written with native 32 bit math, and expects 32 bit numbers when data is passed from one step to another, that's why you would put 32 bit versions of the files and code all together in your release. "That's the precision the worked with from the beginning and what's "expected" or "native" to the program(s). However, you can cast/convert to other data types. Which have different precision (total number of possible values) and obviously 8 bits has less precision than 16, and that's got less than 32. 32 bit float is pretty darn big in terms of possible values--it's fucking huge, but 64 bit or 128 bit numbers as data types. And obviously the bigger data types take more RAM or VRAM. As fenix5fire noted, for "low end" amounts of memory in hardware, casting down to the smaller data type doesn't incur much penalty at all in terms of the result--for this purpose--but takes HALF the amount of RAM or VRAM!
I've been using this for a couple of weeks and it has stopped the constant errors. I looked for quite a while before finding this. Thanks for hosting it on your site.
Thank you there are many areas where access to HF is prohibited, and the download speed is much faster in CIVITAL, and don't worry about the download will be interrupted and lose response
Thanks. I didn't noticed any loading speed increase (idk why i expected this) but still nice to have lightweight one
Man nVidia really screwed us hard this time with their stupid RTX 4090 ridiculous price. I want one.
Does anyone know if the scheduler is available for Forge yet?
is it possible to train a VAE to encode text into images?
Works wonders on my RTX 3050 ^^
it works with qdiffusion?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.