Pipeline properties that read attributes __call__ never sets (SD3 skip_guidance_layers, LEdits++ XL guidance_scale)

[yupengtang]

Describe the bug

A few pipeline properties return self._<name>, but nothing ever assigns that attribute, so reading them raises AttributeError. I checked every @property under src/diffusers that returns a self._x which is never assigned in the class or its bases, and went through the hits by hand (mixins whose subclasses set the value, and helper methods, were false positives). Two real cases are left:

1. StableDiffusion3Pipeline.skip_guidance_layers. __call__ sets _guidance_scale, _skip_layer_guidance_scale, _clip_skip, _joint_attention_kwargs and _interrupt, but not _skip_guidance_layers. Only the property is broken, skip layer guidance itself works. I already opened Fix StableDiffusion3Pipeline.skip_guidance_layers raising AttributeError #14731 with the one line fix, before I had read the updated contribution guide, sorry about the order.

2. LEditsPPPipelineStableDiffusionXL.guidance_scale and do_classifier_free_guidance. Both read self._guidance_scale, which is never set, and __call__ has no guidance_scale argument. __call__ uses them in two places:

   • step 9, when unet.config.time_cond_proj_dim is set (an LCM style UNet): torch.tensor(self.guidance_scale - 1)
   • the ip_adapter_image branch: if self.do_classifier_free_guidance:. That branch already has a # TODO: fix image encoding and fails earlier in encode_image, so this one is mostly moot.

   The first one means LEdits++ XL cannot run with an LCM style UNet at all. I did not want to guess a fix, since LEdits++ has no regular guidance scale in __call__. It could take a new argument, reuse source_guidance_scale from invert(), or reject such UNets up front. Happy to send a PR for whichever you prefer.

Reproduction

LEdits++ XL, using the dummy components from the existing test file with only time_cond_proj_dim set. Run from the repo root:

from tests.pipelines.ledits_pp.test_ledits_pp_stable_diffusion_xl import TestLEditsPPPipelineStableDiffusionXL

t = TestLEditsPPPipelineStableDiffusionXL()
# an LCM style UNet, i.e. unet.config.time_cond_proj_dim is set
pipe = t.pipeline_class(**t.get_dummy_components(time_cond_proj_dim=32))
pipe.set_progress_bar_config(disable=True)

inversion_inputs = t.get_dummy_inversion_inputs()
inversion_inputs["image"] = inversion_inputs["image"][0]
pipe.invert(**inversion_inputs)
pipe(**t.get_dummy_inputs())

With the default components (time_cond_proj_dim=None) the same script runs fine.

For SD3, test_skip_guidance_layers in #14731 reads the property back and fails on main.

Logs

Traceback (most recent call last):
  File "/tmp/diffusers/repro_ledits.py", line 11, in <module>
    pipe(**t.get_dummy_inputs())
    ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/diffusers/src/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py", line 1092, in __call__
    guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
                                         ^^^^^^^^^^^^^^^^^^^
  File "/tmp/diffusers/src/diffusers/configuration_utils.py", line 175, in __getattr__
    raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
AttributeError: 'LEditsPPPipelineStableDiffusionXL' object has no attribute 'guidance_scale'. Did you mean: 'guidance_rescale'?

(the torch decorate_context frame between the first two is left out)

System Info

• diffusers 0.41.0.dev0, main at c419dac
• PyTorch 2.11.0+cu130, Transformers 5.9.0, Python 3.13.15
• Linux, CPU

The reproduction builds tiny random components; the only download is the hf-internal-testing/tiny-random-clip tokenizer.

Who can help?

@yiyixuxu @sayakpaul