[BUG] fixes in kadinsky pipeline

[ishan-modi]

What does this PR do?

Fixes #11060

Who can review?

@DN6

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[yiyixuxu]

thanks @ishan-modi !
are you able to run the docstring examples for these pipelines and see if the outputs are same in this branch vs on main?

[ishan-modi]

I tried running both branches for kadinsky3 and following are the results

they are slightly different potentially due to usage of quantization/balanced_strategy(device_map) because of limited GPU, not sure though

main	fixes-issue-11060	diff

Also preprocess and postprocess are exactly same as before

import torch
import numpy as np
from diffusers.image_processor import VaeImageProcessor
from diffusers.utils import load_image
from diffusers.pipelines.pipeline_utils import numpy_to_pil

input_image = load_image(
 "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky3/t2i.png"
)
midpoint_image = torch.randn(1, 4, 64, 64)
image_processor = VaeImageProcessor(
    vae_scale_factor=2**3,
    vae_latent_channels=4,
    resample="bicubic",
    reducing_gap=1
)

def preprocess(image, image_processor=None, branch="default"):
    if branch == "main":
        arr = np.array(image.convert("RGB"))
        arr = arr.astype(np.float32) / 127.5 - 1
        arr = np.transpose(arr, [2, 0, 1])
        image = torch.from_numpy(arr).unsqueeze(0)
        return image
    
    image = image_processor.preprocess(image)
    return image

def postprocess(image, image_processor=None, branch="default"):
    if branch=="main":
        image = image * 0.5 + 0.5
        image = image.clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        image = numpy_to_pil(image)[0]
        return image
    
    image = image_processor.postprocess(image)[0]
    return image

if torch.equal(preprocess(input_image, image_processor), preprocess(input_image, branch="main")):
    print("Preprocessed images are exactly the same.")

if list(postprocess(midpoint_image, image_processor).getdata()) == list(postprocess(midpoint_image, branch="main").getdata()):
    print("Postprocessed images are exactly the same.")

[yiyixuxu]

thanks!

[yiyixuxu]

I think there are some kandinsky test failure that's potentially related https://github.com/huggingface/diffusers/actions/runs/14330695758/job/40203664079?pr=11080#step:6:33932

[ishan-modi]

Thanks for the review, fixed the tests !

[yiyixuxu]

ohh we can just give them a default value like this

movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) if getattr(self, "movq", None) else ..
movq_latent_channels = self.movq.config.latent_channels if getattr(self, "movq", None) else ..

to be consistent with how it is handled in other pipelines, for example https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py#L218

[hlky]

Same comment for all pipelines.

[hlky]

This breaks output_type == "latent". image_processor.postprocess should not be applied to latent output.

[ishan-modi]

hmm, postprocess function doesn't do anything if the output_type is "latent", see here.

Let me know if I am missing anything

[hlky]

Even so, the code style does not match other pipelines, please make the requested changes to keep consistency with code styling.

[ishan-modi]

Alright, made the change let me know if it looks good

[hlky]

This can be removed. Refer to other pipelines for an example.

diffusers/src/diffusers/pipelines/flux/pipeline_flux.py

Lines 974 to 981 in 31c4f24

	if output_type == "latent":
	image = latents
	else:
	latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
	latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
	image = self.vae.decode(latents, return_dict=False)[0]
	image = self.image_processor.postprocess(image, output_type=output_type)

[ishan-modi]

@hlky gentle ping

[yiyixuxu]

thanks a lot @ishan-modi