After dozens of experiments I’ve settled on the version that’s giving me the most reliable zoom‑outs. I tried training separate LoRAs for extreme, large, and medium zoom levels, but those models were too unpredictable—so I’m sticking with this single “all‑rounder.”Known caveats<ol><li>Out‑paint artifacts – some images pick up chunky/pixelated edges in the newly generated areas.</li><li>Human close‑ups – very tight crops of faces, hands, or bodies can warp badly.</li><li>No precise zoom slider – you can’t tell it “zoom out 10×”; results vary, so experiment.</li></ol>How to prompt Use the base prompt below, then bolt on whatever you’d like to see in the expanded frame. Example:<pre><code>Zoom out and keep the visible subject exactly the same in position, scale and appearance. Expand the canvas evenly in every direction and fill all new areas with a natural continuation of the scene, matching the original lighting, perspective and photographic style. </code></pre>You can change the target latent image size different from your source image size. For example; if your image is vertical, you can expand/apply the zoom out horizontally. Feel free to check the example images. LoRA also works with Nunchaku workflows. That’s it—give it a spin and let me know how it works for you! LoRA trained with <a target="_blank" rel="ugc" href="http://fal.ai">fal.ai</a> Flux Kontext LoRA Trainer, with 70+ pairs, 0.0003 LR, 3000 steps

Files with SHA256: 0aae532bbe5c166d359585dfec983199d8c35158ccfe6ada43594d6ba059ee80