| | --- |
| | license: creativeml-openrail-m |
| | base_model: Qwen/Qwen-Image |
| | tags: |
| | - lora |
| | - qwen-image |
| | - qwen-image-edit |
| | - ai-toolkit |
| | - character-style |
| | - person |
| | --- |
| | |
| | # Model Card: Simone-QI/QIE-v1.0 |
| |
|
| | This is a specialized character LoRA trained as a **Style**, meaning the character is embedded into the model's weights without a trigger word. It is designed for **Qwen-Image** and **Qwen-Image-Edit** (all versions). |
| |
|
| | ## Training Philosophy |
| | The model was trained on 100 high-quality images covering various angles (Front, 3/4, Profile, and Back) and a wide range of outfits and environments. |
| |
|
| | **Caption Strategy:** |
| | * **No Trigger Word:** The character's name, hair color, and eye color were never mentioned in the captions. |
| | * **Full Environment Decoupling:** Everything *except* the character was meticulously tagged (clothing, jewelry, background, lighting). |
| | * **Flexible Generation:** Because the character is the only "unnamed" constant, she appears automatically when you describe a scene. |
| |
|
| | ## Usage & Prompting |
| | Simply describe a woman and her surroundings. The character will manifest as the subject. |
| |
|
| | * **Example Prompt:** `A woman in a red tank top and blue jeans, standing in a rainy forest.` |
| | * **Customization:** Hairstyle can be changed easily at full weight. Eye color is most responsive around 3400-3500 steps (or by slightly lowering LoRA strength). |
| | * **Note:** The model can be sensitive to keywords like "bare" or chest-related descriptions, which may lead to NSFW results. |
| |
|
| | ## Technical Specifications |
| |
|
| | | Parameter | Value | |
| | |---|---| |
| | | **Trigger Word** | None (Trained as Style) | |
| | | **Rank (Dimension)** | 128 | |
| | | **File Size** | 2.2 GB | |
| | | **Training Hardware** | NVIDIA H200 Tensor Core GPU | |
| | | **Training Duration** | ~12 Hours | |
| | | **Total Steps** | 4000 | |
| | | **Bucketing Strategy** | 8 distinct resolution buckets (up to 1536px) | |
| | | **Dataset Quality** | ~8MP source images | |
| | | **Training Toolkit** | Ostris - AI Toolkit | |
| |
|
| | ### Adaptive Batch Size Schedule: |
| | * **0 - 200 steps:** Batch Size 6 |
| | * **200 - 1800 steps:** Batch Size 2 |
| | * **1800 - 2800 steps:** Batch Size 4 |
| | * **2800 - 4000 steps:** Batch Size 10 |
| |
|
| | ### Save Intervals: |
| | * **0 - 3000 steps:** Every 200 steps. |
| | * **3000 - 4000 steps:** Every 100 steps. |
| |
|
| | ## Sample Images (Upscaled) |
| | Below are samples showcasing the character's consistency across different outfits, angles, and environments. |
| |
|
| | ## Sample Images |
| | Below are samples showcasing the character's consistency across different outfits, angles, and environments. |
| |
|
| | ### Portrait Gallery |
| | | | | |
| | |:---:|:---:| |
| | |  |  | |
| | ### Wide Perspective |
| | | | |
| | |:---:| |
| |  |
| | ### Focus & Details |
| | | | | |
| | |:---:|:---:| |
| | |  |  | |
| | |  |  | |
| |
|
| |
|
| | --- |
| | *Model card generated for the Simone-QI series.* |
