A side-project of mine, due to my love of outer space and the greater universe.

A bit of an oversight, I suppose, space happens to be. Most models produce honestly garbage or 3D computer generated planets, space imagery/videos. I wanted a real NASA style, actual earth footage, but generatable as a video.

A video LoRA for MiniMax H3 focused on realistic views of Earth from low orbit: curved horizons, a thin blue atmosphere, deep blue oceans, detailed cloud formations, and desert coastlines against black space, with spacecraft structures framing the view. Designed for ISS-style orbital footage and cinematic planetary shots.

Example prompt:
"This is a high-resolution photograph of Earth from space, taken from a satellite or spacecraft. The image prominently features the curved horizon of the planet, showcasing a vivid blue ocean with scattered white clouds. The ocean occupies the lower half of the image, while the upper half is dominated by the cloud-covered atmosphere. The clouds are dense and fluffy, creating a textured pattern against the bright blue of the ocean. In the top right corner, parts of the spacecraft's structure are visible, including metallic components and a section of the satellite's arm. The background is the blackness of outer space, providing a stark contrast to the bright colors of Earth. The photograph captures the beauty and vastness of our planet, highlighting the interplay between the ocean and clouds. The overall composition emphasizes the curvature of Earth, making the planet appear almost spherical. The image is clear and detailed, with a sharp focus on the Earth's surface and a slightly blurred background due to the depth of field."

If you like it, if you happen to use it often; consider donating gold buzz for my future projects, and improvements to my current ones. Many people don't realize how expensive renting datacenter GPUs can be, and how trial and error, even at $2 an hour, can quickly become $50, $100, or even $200. Locally, I am only on an RTX 4070 Ti Super, limiting the VRAM to 16GB definitely makes video training certainly difficult if not impossible at higher resolution buckets, with longer video clips. That, plus the cost of $100 Chatgpt Pro, which has, with codex, been my go-to for analyzing loss charts, lora training failures, safetensors themselves by measuring tensor values/floating points, and predicting safer-more reliable training data by sorting much of it, before I get to.

Just letting all of you who maybe are not familiar with LoRA creation/AI image/video training know the breakdown. It's time consuming, and costly if local hardware is a factored limitation. (sometimes is). I can manage costs, but I am not willing to spend extra past a certain threshold, unless someone wants to fund this.

I am open to LoRA creation if you want anything specific, broad, or wildly experimental made; contact me. If you throw funds at me, you will have quality LoRA! (I am persistent, and will not stop until it is EYE-CANDY level.