[Bug]: ADetailer is not working in Forge Neo when using z-image

[tottem]

Describe the bug

Hello.
ADetailer detects the face perfectly but does not apply the transformation. I created a Lora of a character, basically just the face. I don't like applying the Lora during image generation, because even though the character's resemblance is quite accurate, applying the Lora always distorts the image, especially the rest of the character's body. Because of this, I always prefer to create the image without the Lora and then apply the Lora in ADetailer, which usually gives me a good image without distortions and a strong resemblance to the character. However, when using z-image, ADetailer in Forge Neo does not complete the job

Steps to reproduce

No response

Screenshots

No response

Console logs, from start to end.

No console log

List of installed extensions

No response

[dosubot]

Hi @tottem! I'm Dosu and I’m helping the adetailer team.

This is a known limitation with how Forge Neo integrates ADetailer: when ADetailer runs in Forge Neo (especially after using z-image), it resets the active scripts and only reinserts ControlNet, dropping any other custom scripts—including LoRA handlers. As a result, any LoRA layers applied during the z-image step are not carried over into ADetailer’s processing, so the transformation is not applied even though detection works. There’s no explicit z-image or LoRA handling in ADetailer, and this script reset is the main reason for the missing transformation you’re seeing in Forge Neo with z-image workflows (details).

Possible workarounds: try applying the LoRA during the initial generation (with prompt tuning to minimize distortion), use a different workflow/platform where script chaining is preserved, or manually edit the Forge Neo pipeline to retain LoRA scripts in the post-processing step.

If you can share your exact versions and workflow steps, it might help identify more targeted solutions.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[tottem]

Thanks Dosu, so the issue is forge neo. Do you know of any other fork, a reforge perhaps, that can handle adetailer using z-image?

[dosubot]

Right now, there isn’t any fork—including ReForge—that’s confirmed to handle ADetailer with z-image workflows and preserve LoRA/script chaining. The main issue is that Forge Neo (and similar forks) reset the active scripts and only reinserts ControlNet, dropping custom scripts like LoRA handlers, so transformations don’t apply after z-image details.

ReForge does support ADetailer and works with certain ultralytics versions for segmentation models, but it has its own compatibility problems and there’s no evidence it fixes the script reset or LoRA chaining issue with z-image workflows discussion. Most issues in ReForge are related to ultralytics integration, not script chaining.

If you need LoRA to apply post-generation, the only reliable method is to use a platform where script chaining is preserved, or manually modify the pipeline to keep LoRA scripts active during ADetailer runs.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[tottem]

Ok. Thank you very much...last question and thanx for your time. What about comfyUi.... Do you know if adetailer could be integrated in a node in comfyUi.

[dosubot]

There isn't an official or community-supported ADetailer node for ComfyUI right now. The main reason is that ADetailer is tightly integrated with AUTOMATIC1111's WebUI and uses script callbacks and UI hooks that aren't compatible with ComfyUI's node system. There is community interest in a ComfyUI port, but no one has released a plugin or node yet discussion.

If you're looking for similar functionality in ComfyUI, check out nodes like the SEGM detector from ComfyUI-Impact-Pack, which offers segmentation and class filtering features. However, these aren't direct replacements for ADetailer—they cover some, but not all, of its features feature request. Porting ADetailer to ComfyUI would require significant development work and isn't currently in progress.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[Dawgmastah]

Am I seeing things? I do see the effect of adetailer on the z-Image on Forge neo, (If the initial image already has the lora applied)

Wouldnt it be feasible then to do an initial inference with the lora lora:.01 then on adetailer specifically scale to lora:1?

[Dawgmastah]

Actually I just confirmed this works fine, one can even insert a completely different lora, not used in either normal res or hires, and the adetailer result will work as expected.