Sign In
homemodelsimagesvideospostsarticlesbountieschallengeseventsupdatesshop

ComfyUI Updates Changed My Z-Image Turbo LoRA Output (and How I Fixed It)

6

Ggrue's Avatar

Ggrue

Supporter Badge March 2024

Jan 6, 2026

(Updated: 2 months ago)

training guide
zimagebugcomfyuitraininglora

Problem

After updating ComfyUI to versions 0.3.76 or more (in my case 0.4.0), my Z-Image Turbo LoRA started producing different images even with:

  • identical workflow

  • identical prompts

  • identical seed

  • identical LoRA strength settings
    (only ComfyUI was different)

z_00047_.png

Before Comfy update

z_00475_.png

Afer Comfy update

Typical symptoms:

  • the LoRA looks “too strong” or “wrong”, the details are lost.

What’s Going On

Z-Image Turbo LoRAs target specific internal modules (layers) of the model. When ComfyUI updates how Z-Image Turbo LoRA layers are mapped/loaded/applied, a LoRA that was tuned for an older mapping can suddenly affect a different set of submodules (or affect more of them), which changes the output.

In practice, the most sensitive part tends to be the attention projections:

  • to_q

  • to_k

  • to_v

  • to_out

If ComfyUI starts applying these attention LoRA layers differently than before, the “same LoRA at the same weight” can behave very differently.

The Fix: Remove Attention LoRA Layers (Pruned / Compatibility Variant)

The workaround that gave me stable behavior again was to make a “compatibility/pruned” variant of the LoRA that removes the attention projection LoRA weights and keeps the rest (often MLP/FFN/etc.).

This has two practical advantages:

  1. Stability: it avoids the most version-sensitive part of the LoRA application.

  2. Smaller file size: removing those tensors reduces the LoRA size significantly.

Conceptually:

  • “Muting” a layer (making it have no effect) can be done either by setting weights to zero or by removing those weights entirely.

  • Removing them is cleanerand gives a smaller file.

Same image with the pruned version of the Lora:

z_00479_.png

As you can see it's almost identical to the last version of Comfy !

I link you the python script I used in the attachments. Good news is that it's almost instant to process so no need to retrain.

To use it: place it in your lora folder then in cmd type:

python prune_zit_attention.py your_lora.safetensors your_lora_pruned.safetensors

Can we improve it further ?

Yes by using the irght sampler and scheduler when generating il will improve the details, skin texture and the lora effect without degrading the image like when increasing the lora weight.

You should be using
Sampler dpmpp_2s_ancestral

Scheduler FlowMatchEulerDiscreteScheduler

"FlowMatchEulerDiscrete seems not exposed in ComfyUI, but it is what the official Z-Image demo in diffusers use.

In short...if you want sharper and noise free images, use this!"

If that scheduler does not appear in your scheduler list, you need to install it as a custom node from https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler
git clone https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler.git in your custom nodes folder.

Resulting image for us (keeping all other parameters the same)

z_00482_.png

Let me know if this guide helps you or if you have any question !

You can find my loras on my profile and my celebrities loras on my own website.

All my links here.

Thanks to alcaitiff who made me understand this issue by making a pruned version of their lora

6