Skip to content

🟢[SDXL FEATURE REQ] EQ VAEs and padding patch [sd3 branch] #2189

@deGENERATIVE-SQUAD

Description

@deGENERATIVE-SQUAD

We have two branches of new, especially precise and “clean” VAEs for SDXL called EQ VAE (one from https://huggingface.co/KBlueLeaf/EQ-SDXL-VAE ). Test trainings show a visually noticeable improvement when using these new VAEs for training.

Problems:

  1. Anzhc’s EQ VAE (https://huggingface.co/Anzhc/MS-LC-EQ-D-VR_VAE) loads correctly, but also requires the patch from the project page to dodge the padding bug:
for module in self.model.modules():
            if isinstance(module, nn.Conv2d):
                pad_h, pad_w = module.padding if isinstance(module.padding, tuple) else (module.padding, module.padding)
                if pad_h > 0 or pad_w > 0:
                    module.padding_mode = "reflect"

Additional module for the inference fix via Comfy with models trained on Anzhc’s EQ VAE here https://github.com/Jelosus2/comfyui-vae-reflection

It would be great to be able to use in the full extent, if that’s not too much trouble to implement.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions