V2.1 Update: Expanded Dataset &amp; Better Generalization<ul><li>Dataset Expansion: The dataset has increased from 159 to 350 images, covering a wider range of styles and scenarios. Approximately 40% are synthetic data generated using the previous version (v1.0) or Grok Imagine.</li><li>Quality &amp; Style: Significantly reduced the negative impact on image quality (less burning). The model now works with a broader variety of styles, though it still performs best with anime aesthetics.</li><li>Flexibility vs. Control: The model is now much more creative and can generate stabbing scenarios with various conditions, poses, and objects—even those not explicitly found in the dataset. However, it is harder to control than previous version, prompt adherence is lower when using complex combinations of triggers.</li><li>Usage Tip: For best results, place the trigger words at the very beginning of your prompt.</li></ul>The usage guide has also been updated for v2.1.<hr />Model DescriptionThis LoRA is designed to add toy-like decorations (resembling stabbed objects) attached to characters at various positions. It provides control over the object type, the wielder, and the angle. Please experiment creatively!⚠️ Disclaimer: I am not responsible for any inappropriate content created using this model.Recommended Settings<ul><li>Style: Works best with Anime styles.</li><li>Checkpoints: Tested primarily with Wai-illustrious and Wai-Rouwei.<ul><li>Note: The Wai-Rouwei model tends to have a higher success rate for complex compositions.</li></ul></li><li>LoRA Weight: 0.4 ~ 0.8.<ul><li>Higher weights improve prompt adherence but may degrade image quality.</li><li>Adjustment: You may need to retune the weight when switching base models.</li></ul></li><li>Resolution: Higher resolutions work better, especially for larger "toys," as this LoRA was originally trained for inpainting.</li></ul>💡 Pro Tip for High Quality: Use a higher weight first to get the correct composition. Then, use that image as input for Depth ControlNet and regenerate with a lower LoRA weight to polish the details.Trigger Words1. Position Control (Mutually Exclusive) Adding "[object] into [body part]" to your prompt also improves accuracy.<ul><li><code>abdomen_stabbed</code></li><li><code>chest_stabbed</code></li><li><code>head_stabbed</code></li><li><code>neck_stabbed</code> (Weakest effect due to limited training data)</li><li><code>stabbed in the back</code> (Limited viewing angles)</li></ul>2. Wielder &amp; Scenario Control<ul><li><code>independent_stabbed</code>: (Default) No wielder. Use this if you have multiple characters but want the object to appear unsupported.</li><li><code>assisted_stabbed</code>: Use this if someone else is holding the object.<ul><li>Tip: This requires more canvas space. Add <code>solo focus</code> to prioritize the stabbed character, otherwise, the composition may split focus evenly. Prompts like "confrontation" or "fight" help set the atmosphere.</li></ul></li><li><code>impaling_stabbed</code>: Shows both entry and exit points.<ul><li>Tip: Success depends heavily on the viewing angle and object length (e.g., "long sword" works better than "dagger").</li></ul></li></ul>3. Viewing Angle(optional) I carefully tagged different angles during training. Specifying the correct angle can help for a logical composition.<ul><li><code>front view</code>, <code>from side</code>, <code>3/4 view</code>, <code>back view</code>, <code>over-the-shoulder shot</code>, <code>wide shot</code>, <code>high angle</code>, <code>low angle</code>, <code>from above</code>, <code>from below</code>, <code>POV</code></li><li>Avoid: Contradictory prompts (e.g., "front view" + "stabbed in the back")</li></ul><hr /><hr />Prompting Structure Template (Updated for V2.1)Here is a recommended structure to get the best results:Structure: <code>[Main Trigger], [Optional Trigger], [Involved Character], [Object Name], [Pose], [Location Fine-tune], [Viewing Angle], BREAK, [Character Name], [Outfit/Description], [Background], [Quality/Style Tags]</code>Example: <code>abdomen_stabbed, self_stabbed, 1girl, solo, toy sword, standing,holding sword, toy sword into stomach, from side, upper body, BREAK, hatsune miku, smiling, looking at viewer, teal twintails, skirt, futuristic city, masterpiece, best quality, anime screencap</code>💡 SDXL Pro Tip: Using the <code>BREAK</code> command to separate prompt segments is highly recommended for SDXL models to improve concept adherence.<ul><li>Note for ComfyUI Users: You generally need to install a custom node (e.g., via nodes that support prompt styling) for the <code>BREAK</code> command to function correctly.</li></ul>Negative Prompt Guide Since "stab" is a concept often associated with violence in base checkpoints, strict negative prompting is essential if you want to avoid inappropriate outcomes.<ul><li>Essential Tags: <code>blood, guro, nsfw</code> (Prevents inappropriate gore generation).</li><li>Quality Tags: I recommend using negative embeddings (e.g., <code>lazyneg</code>) to reduce anatomical errors.</li><li>Object Control: <code>multiple weapons, multiple [Object Name]</code><ul><li>helps prevent the model from generating duplicate objects or "ghost" weapons in the scene.</li></ul></li></ul>

stabby toy sword

Dataset: Increased to 350 images (40% synthetic data via v1.0 &amp; Grok).Quality: Significantly reduced impact on base model quality; supports more art styles (anime still recommended).Generalization: Much higher creativity. Can generate scenes/poses/objects not present in the original dataset.Usage: Control is slightly more sensitive. Move main trigger words to the beginning of prompts.

Files with SHA256: 47001fc8c5ffebd7037e181c74a2e8e7e5677117acc98fb29590b099e676508a