<h1 id="sinistersmile">😈 SinisterSmile</h1><h3 id="a-mean-looking-happy-face-bobblehead-mask-lora-for-krea-2-z-image-flux.2-klein-9b-hidream-and-qwen-image">A Mean-Looking Happy Face Bobblehead Mask LoRA for Krea 2, Z-Image, Flux.2 Klein 9b, HiDream &amp; Qwen Image</h3>SinisterSmile transforms ordinary portraits into unsettling, stylised characters wearing an oversized bobblehead mask featuring an exaggerated happy face with an unmistakably sinister edge. The contrast between the cheerful grin and threatening expression creates imagery that feels playful, eerie and instantly memorable.Trained specifically for Krea and Qwen Image, this LoRA excels at generating highly detailed mascot-style masks, oversized heads, smooth plastic finishes and expressive facial features while maintaining strong prompt adherence. Whether you're producing horror-inspired artwork, liminal scenes, surreal portraits, streetwear concepts or creepy internet aesthetics, SinisterSmile delivers a distinctive look that stands out.The model performs particularly well when combined with cinematic lighting, abandoned environments, colorful urban settings and dramatic compositions, making it equally suitable for horror projects and stylized digital art.<hr /><h2 id="trigger-word">Trigger Word</h2><code>smiley</code><hr /><h2 id="recommended-settings">Recommended Settings</h2>Weight: 0.8–1.0Lower weights create subtle mascot-inspired features, while higher values emphasize the oversized bobblehead proportions, sculpted grin and exaggerated facial details.Flux Only: 0.4 - 0.6Flux tends to be more powerful so loras don't need the high strength level.<hr /><h2 id="example-prompt">Example Prompt</h2><pre><code>smiley, person wearing a sinister happy face bobblehead mask, oversized mascot head, creepy smile, glossy plastic mask, cinematic lighting, abandoned hallway, dramatic shadows, ultra detailed, realistic textures, masterpiece</code></pre><hr /><h2 id="description">Description</h2>The appeal of SinisterSmile lies in contradiction. A smiling face usually conveys comfort and friendliness, yet this oversized mask feels strangely unsettling. That tension creates images that instantly capture attention, whether used for horror concepts, album artwork, YouTube thumbnails, social media graphics or AI-generated storytelling.Because the LoRA was trained to preserve the distinctive proportions and expression of the mask rather than a single scene, it adapts well to a wide variety of prompts. Place the character in a neon-lit alley, a dark forest, a shopping mall or a bright studio and the signature grin remains recognisable without overwhelming the rest of the composition.<hr /><h2 id="best-uses">Best Uses</h2><ul><li>Horror artwork</li><li>Mascot characters</li><li>Creepy portraits</li><li>Album covers</li><li>YouTube thumbnails</li><li>Streetwear concepts</li><li>Surreal photography</li><li>Horror game concepts</li><li>Social media artwork</li><li>AI storytelling</li></ul><hr /><h2 id="tips">Tips</h2>For the strongest results, combine SinisterSmile with cinematic lighting, volumetric fog, dramatic colour grading or shallow depth of field. Pairing it with expressions such as standing still, staring at the viewer, walking toward camera or head tilted often enhances the unsettling atmosphere.<hr /><h2 id="if-you-enjoy-this-lora">If You Enjoy This LoRA</h2>If you enjoy using SinisterSmile, consider leaving a ⭐ rating and sharing your creations. Seeing what the community creates helps improve future releases and inspires new LoRA concepts.If you'd like to learn how models like this are created from scratch, check out my complete LoRA Training Guide, where I walk through the entire workflow—from dataset preparation and captioning to training, optimisation and publishing. Whether you're new to AI model creation or looking to refine your process, the guide is designed to help you produce higher-quality LoRAs with confidence.Happy creating!Note:Thank you modelscope and diffsynth for the training scripts

SinisterSmile

Trained on Qwen Image 2152Krea 2 Version: <a rel="ugc" href="https://www.kofi.com/tensorvizion">https://www.kofi.com/tensorvizion</a>

SmileGang-Qwen_c1-st3000.safetensors

Qwen 2152

Animagine-XL

Illustrious

Flux.2 Klein 4B

Flux.2 Klein 9B

HiDream


# 

## Model Introduction

This model is trained using the [Civision/model training](https://modelscope.ai/civision/modelTraining) environment and computing resources provided by ModelScope.

* Model Type: LoRA
* Base Model: [Qwen/Qwen-Image-2512](https://modelscope.ai/models/Qwen/Qwen-Image-2512)
* Training Code: [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio)
* Data amount：14
* Total training steps：3000
* License: Apache-2.0

## Inference Code

Install [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio):

```bash
pip install diffsynth
```

Inference:

```python
from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig
import torch

pipe = QwenImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="Qwen/Qwen-Image-2512", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors"),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="text_encoder/model*.safetensors"),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="tokenizer/"),
)
pipe.load_lora(pipe.dit, ModelConfig(model_id="tensorvizion/SmileGang-Qwen", origin_file_pattern="SmileGang-Qwen_c1-st3000.safetensors"))
prompt = "a cat"
image = pipe(prompt)
image.save("image.jpg")
```


SmileGang-Qwen

c1-st3000

c1-st1000

c1-st2000

Files with SHA256: 4dbc8f893abc89bed300bd30dd55a8de1b774e89146939da93cda9c1226155bb