---
license: apache-2.0
base_model: Tongyi-MAI/Z-Image-Turbo
tags:
  - lokr
  - lora
  - z-image
  - repro
---

# Z-Image LoKr — diffusers loading repro

**Random weights — not a functional adapter.** This file reproduces the key layout of a real
Z-Image LoKr trained with an ai-toolkit/LyCORIS-style trainer (ComfyUI `diffusion_model.` key
format), for reproducing https://github.com/huggingface/diffusers/issues/13221.

`ZImagePipeline.load_lora_weights` / `lora_state_dict` fails on it with
`ValueError: state_dict should be empty at this point` because
`_convert_non_diffusers_z_image_lora_to_diffusers` has no handling for `lokr_w1` / `lokr_w2` keys.

```python
from diffusers import ZImagePipeline
from huggingface_hub import hf_hub_download

path = hf_hub_download("scenario-labs/z-image-lokr-repro", "zimage_lokr_dummy.safetensors")
ZImagePipeline.lora_state_dict(path)  # raises ValueError on diffusers <= 0.39 / main
```

## Layout (612 tensors, 30 layers)

Mixed LoKr + plain LoRA:

| module | format | layers | shapes |
| --- | --- | --- | --- |
| `attention.to_q/k/v` | LoKr | all 30 | `lokr_w1 [4,4]`, `lokr_w2 [960,960]`, scalar `alpha` |
| `adaLN_modulation.0` | LoKr | all 30 | `lokr_w1 [4,4]`, `lokr_w2 [3840,64]`, scalar `alpha` |
| `attention.to_out.0` | LoRA | 0–17 | `lora_A [32,3840]`, `lora_B [3840,32]` |
| `attention.to_out.0` | LoKr | 18–29 | `lokr_w1 [4,4]`, `lokr_w2 [960,960]`, scalar `alpha` |
| `feed_forward.w1/w2/w3` | LoRA | all 30 | rank 32 |

The same file loads fine in current ComfyUI (which fuses `to_q/k/v` into `attention.qkv` and
supports LoKr via `comfy/weight_adapter/lokr.py`).