TangentialClassifierFreeGuidance.is_conditional reads an attribute that does not exist

[ojassharma7]

Describe the bug

TangentialClassifierFreeGuidance.is_conditional reads self._num_outputs_prepared, which is not defined on the class, on BaseGuidance, or anywhere else in the package. Accessing it raises AttributeError, and so does is_unconditional, since BaseGuidance defines that as not self.is_conditional.

BaseGuidance.__init__ sets self._count_prepared, and prepare_inputs increments it. Every other guider reads that attribute. This one appears to be the only exception:

adaptive_projected_guidance.py        return self._count_prepared == 1
adaptive_projected_guidance_mix.py    return self._count_prepared == 1
auto_guidance.py                      return self._count_prepared == 1
classifier_free_guidance.py           return self._count_prepared == 1
classifier_free_zero_star_guidance.py return self._count_prepared == 1
frequency_decoupled_guidance.py       return self._count_prepared == 1
magnitude_aware_guidance.py           return self._count_prepared == 1
perturbed_attention_guidance.py       return self._count_prepared == 1 or self._count_prepared == 3
skip_layer_guidance.py                return self._count_prepared == 1 or self._count_prepared == 3
smoothed_energy_guidance.py           return self._count_prepared == 1 or self._count_prepared == 3
tangential_classifier_free_guidance.py  return self._num_outputs_prepared == 1   <-- undefined

grep -rn "_num_outputs_prepared" src/ returns that single line and no assignment.

The reported error is misleading, which is what makes this awkward to diagnose. Python surfaces a failure inside a property as the property itself being missing, so the message names is_conditional rather than the attribute that is actually absent.

The property is reachable rather than dead code. BaseGuidance.is_unconditional routes through it, and modular_pipelines/stable_diffusion_xl/denoise.py reads components.guider.is_conditional in the guess-mode branch.

Reproduction

No GPU, model download or authentication required.

from diffusers.guiders import ClassifierFreeGuidance, TangentialClassifierFreeGuidance

print(ClassifierFreeGuidance(guidance_scale=7.5).is_conditional)          # False

t = TangentialClassifierFreeGuidance(guidance_scale=7.5)
print(hasattr(t, "_count_prepared"))          # True
print(hasattr(t, "_num_outputs_prepared"))    # False
t.is_conditional                              # AttributeError
t.is_unconditional                            # AttributeError

Logs

AttributeError: 'TangentialClassifierFreeGuidance' object has no attribute 'is_conditional'

Prior attempts

Two pull requests fixed this correctly and are now closed, in both cases by their own authors rather than by a maintainer or a review decision:

• fix: correct AttributeError in TangentialCFG and NameError in FrequencyDecoupledGuidance #13434 (2026-04-08, closed 2026-04-27 by the author) also covered a NameError in FrequencyDecoupledGuidance. That second bug no longer reproduces on main, so only this one remains.
• guiders/tangential_classifier_free_guidance: fix AttributeError in is_conditional property #13764 (2026-05-19, closed 2026-07-21 by the author) received no comments at all.

I mention this because it seems more useful than filing a third identical patch unprompted. Given that guiders are marked experimental, is this worth fixing, or is the module parked for now? If you would like it fixed I am happy to open a PR changing the attribute to _count_prepared, with a test that covers is_conditional and is_unconditional for every guider so the next one cannot drift the same way.

System Info

• diffusers 0.41.0.dev0, main at ae2e4c7
• torch 2.14.0, Python 3.11, macOS (Apple Silicon)
• Reproduced on CPU, no accelerator involved

Who can help?

@DN6 @asomoza