I would encourage you to post your generated images here if it's working. A larger dataset will help refine the LoRA or train the concept for another base model.<hr />It generates a horizontally positioned subject being carried by the viewer, with high modification flexibility through prompting.Depending on your framing, the body proportions and camera perspective might look weird sometimes, which is totally expected because there's no 2D camera lens capable of replicating the true binocular first-person vision.Example Prompt:<pre><code>povBriCarry, POV first person view, looking down at [] being carried.
[]'s head and body is positioned on the [] side of frame, lying horizontally with legs extending to the [], wearing [].
The viewer is supporting the [] with [chest/arm] visible, wearing [].
The background is [].</code></pre>Recommended Weight: 0.5 to 1.0Works best with the official Krea2 base model. Known overfitting issues: the subject's arm tends to reach towards the viewer's shoulder, and sometimes it generates jewelry without mentioning explicitly. Use negative prompt to prevent them.<hr />This is probably the hardest LoRA I've ever attempted. There's absolutely zero existing image I can find to build the dataset.Those anime style concept LoRAs on this site only produce anatomically broken and perspective distorted images, which are most likely unusable if I want to maintain Krea's realistic agility. And even some of the SOTA closed-source models (I used Gemini) can't understand the complex spatial logic of this pose from text prompts alone, so I had to build reference manually with 3D models and use controlnet to generate a synthetic dataset, but even controlnet struggled with it. I only managed to generate over a dozen high-quality images after countless iterations and manual repainting. It was really a troublesome process.

Files with SHA256: e4ec0904afffd66f9db74b2c4fcc554b8bd62e18678c587715eea33185081dd4