[bug]: 3.2.0 - Latent to Image crashes when using TAESD

invoke-ai / InvokeAI

InvokeAI is a leading creative engine for Stable Diffusion models, empowering professionals, artists, and enthusiasts to generate and create visual media using the latest AI-driven technologies. The solution offers an industry leading WebUI, supports terminal use through a CLI, and serves as the foundation for multiple commercial products.

https://invoke-ai.github.io/InvokeAI/

Apache License 2.0

22.31k stars 2.32k forks source link

[bug]: 3.2.0 - Latent to Image crashes when using TAESD #4722

Open grunblatt-git opened 9 months ago

grunblatt-git commented 9 months ago

Is there an existing issue for this?

[X] I have searched the existing issues

OS

macOS

GPU

mps

VRAM

What version did you experience this issue on?

3.2.0

What happened?

When using TAESD as VAE, the image generation crashes with the following message

[InvokeAI]::ERROR --> Error while invoking: 'DecoderTiny' object has no attribute 'mid_block'

Screenshots

No response

Additional context

This seems to happen independently from the main model used. The same error appears when invoking from the main UI and from within workflows.

TAESD was installed via the Model Manager by providing the link madebyollin/taesd as suggested in the Release Notes

Contact Details

No response

zealsprince commented 9 months ago

Seconding this as I'm encountering the same issue. I don't believe it should be used as the VAE though unless I am mistaken. I was excited for this feature but it is not fully implemented as per #4056

grunblatt-git commented 9 months ago

Maybe i am just using it wrong, but my understanding from the TAESD documentation was that it is used instead of a usual VAE. InvokeAI also classifies the model as VAE, so using it instead of another VAE seemed intuitive.

TAESD is very tiny autoencoder which uses the same "latent API" as Stable Diffusion's VAE.

The Release Notes mention that TAESD is supported, but as you pointed out using TAESD for live-preview is still work in progress (see #4056). So i assumed that using it as VAE is the only way TAESD is (or should be) supported as of v3.2.0.

Millu commented 9 months ago

Yep TAESD is used as a VAE - it stands for Tiny Auto Encoder for Stable Diffusion

~@grunblatt-git I'm not having any issues using it, these are my generation settings on 3.2.0rc3:~

I'm now experiencing the same issue

Millu commented 9 months ago

Update - it should work using fp16/half precision