Workflow's prompt adherence is significantly weaker than ModelScope's official WenbUI in Qwen image Edit

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

#### 1. Expected Behavior (Based on ModelScope Website)

* **High Prompt Adherence:** The model should be able to significantly alter the image content and composition based on text prompts (e.g., "spring" at least water in my test).
* **High Batch Diversity:** When `batch` is set to 4, the official site produces 4 images with distinct differences in composition and style.
* **Performance Benchmark:** [ModelScope Official Webui](https://modelscope.cn/aigc/imageGeneration?imageId=18281703&tab=advanced) (Please see the generation results at this link; it clearly adheres much better to the prompt).

### Actual Behavior

#### 2. Actual Behavior (ComfyUI Node)

* **Low Prompt Adherence:** The node generates images that are heavily "stuck" to the original style and composition of the reference image.
* **Prompt Ineffectiveness:** It seems only the very first few prompts have a effect. Prompts further down the list (like "spring") **are almost completely ignored**.
* **Low Batch Diversity:** When `batch` is set to 4, the ComfyUI node produces 4 images that are nearly identical, lacking any meaningful variation.

[Qwen-Image-Edit_2509.json](https://github.com/user-attachments/files/23587846/Qwen-Image-Edit_2509.json)

I also found that the seed on ModelScope is within the int range, whereas KSampler's is much higher than this. I'm not sure if this is related.

### Steps to Reproduce

**[Environment and Settings]**

* **Runtime Environment:** RuningHub cloud platform
* **Parameters:** I have done my best to align parameters in the Workflows (e.g., Steps, CFG, Seed) with those on the ModelScope website.
* **Quantization:** Confirmed that no quantization was used.
* **My Workflow:** [Qwen-Image-Edit_2509_single_image_fusion.json](https://github.com/user-attachments/files/23587586/Qwen-Image-Edit_2509.json)

### Debug Logs

```powershell
the cloud platform doesn't provide terminal access, and I am unable to run the full Qwen Image Edit model locally
```

### Other

**[Image Evidence]**

* **Reference Image:**

* **ModelScope (Expected Result:**
* As shown in the two images below, although the character's art style differs more from the reference, it **is clearly trying to follow the text prompt**.
* With `batch=4`, the composition changes significantly, **successfully incorporating later prompts like "spring"**.

* **ComfyUI (Actual Result):**
* As shown in the two images below, it **completely fails to incorporate prompts like "spring"** and just rigidly reproduces the reference image's style.

* This level of low prompt adherence is comparable to my local 8-step LoRA + fp4 svdq quantized version (though the 8-step LoRA has even worse diversity, of course).