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
Out‑paint artifacts – some images pick up chunky/pixelated edges in the newly generated areas.
Human close‑ups – very tight crops of faces, hands, or bodies can warp badly.
No precise zoom slider – you can’t tell it “zoom out 10×”; results vary, so experiment.
How to prompt
Use the base prompt below, then bolt on whatever you’d like to see in the expanded frame. Example:
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.
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 fal.ai Flux Kontext LoRA Trainer, with 70+ pairs, 0.0003 LR, 3000 steps
