Wrong Z_Image guidance scale implementation

[Trgtuan10]

Describe the bug

With cfg implementation of Z_image: pred = pos + current_guidance_scale * (pos - neg), the do_classifier_free_guidance function should be implement: return self._guidance_scale > 0, but currently in diffusers code: return self._guidance_scale > 1

For example, if I set guidance = 0.5, it works as guidance = 0 and it is incorrect.

Reproduction

No

Logs

System Info

0.36.0

Who can help?

@yiyixuxu @sayakpaul @DN6

[Trgtuan10]

Sorry if it shouldn't be in bug category

[sayakpaul]

Cc: @RuoyiDu @JerryWu-code from Z-Image team. They are the ones that contributed the pipeline.

[asomoza]

@Trgtuan10 this is not exclusive to Z-Image but to all diffusers, see for example here

The only part that is missing is the text with the explanation:

here guidance_scale is defined analog to the guidance weight w of equation (2)
of the Imagen paper: https://huggingface.co/papers/2205.11487 . guidance_scale = 1
corresponds to doing no classifier free guidance.

Probably you're coming from the UIs where you use 0 for no guidance, but in diffusers all the models share the same logic which is the correct implementation of the paper (but I understand that it can be confusing in the UX perspective)

[Trgtuan10]

Thanks for the clarification — I understand the diffusers convention that guidance_scale = 1 means no CFG . That said, I think Z-Image is currently using a different CFG composition than the rest of diffusers:

• Most pipelines do: pred = neg + guidance_scale * (pos - neg) so when guidance_scale = 1, pred = pos (no guidance), and when guidance_scale <= 1, it’s still well-defined.
• Z-Image does: pred = pos + guidance_scale * (pos - neg) which makes guidance_scale = 0 correspond to pred = pos (no guidance), and guidance_scale = 1 correspond to pred = 2*pos - neg.

So with the current do_classifier_free_guidance check (> 1), setting guidance_scale = 0.5 won’t trigger CFG, even though the formula would still apply meaningful scaling.

Would it make sense to re-implement Z-Image to follow the same convention as other pipelines (neg + w*(pos-neg)), so the guidance_scale semantics and the do_classifier_free_guidance logic are consistent across diffusers? If so, I’m happy to open a PR.

[asomoza]

oh I didn't look at the rest of the code, I was just referencing to the line you linked, yeah, it's different but the reason as to why is like that and the decision to change it depends on the model authors that were cced before, maybe they have a good reason as to why is different, the same as the dynamic shifting code that's not used.

[Trgtuan10]

Yeah, maybe @RuoyiDu @JerryWu-code can answer ...

[yiyixuxu]

cc @JerryWu-code here
I think @Trgtuan10 is correct and we should change

do_classifier_free_guidance to return self._guidance_scale > 0

context is diffusers implement CFG defined in Imagen paper, Z-image seems to use the format from the original CFG paper

[Trgtuan10]

Should I create a PR for it @yiyixuxu ?