I was watching a Sci Fi show and saw this one scene that made me wonder.... Can I train a LORA from a single 5 second scene? The answer: "sure! why not!?"<h1 id="stuck-in-a-gelatinous-blob">Stuck in a Gelatinous Blob</h1>Don't ask why... ask 'how'! We'll get to the why when the blob is done and spoiler alert, the blob is never done.Since I only had a single scene to work with, I varied the captions, flipped the videos horizontally, shifted the colors and played with cropping as well. The result, a surprisingly not-useless LORA!<h3 id="now-in-minimax-h3!">Now in MiniMax H3!</h3>Wildcard Prompt<pre><code>A beautiful {Russian|French|Swedish|Swiss|African|Latina|Austrian|Dutch|English|American|Siberian|Chinese|Ukrainian|Thai|Czech} woman is stuck in a {green|clear|yellow|red|blue|pink|white|black|multicolor} gelatinous blob, with her head and shoulders visible above the surface.

She has {very long|long|shoulder length} {blonde|brown|dirty blonde|black|auburn} hair.

She is {screaming from pleasure|very surprised and shocked|furrowing her brow and gritting her teeth|throwing her head back}.

They are {inside a tent with the door flap open behind them showing {snowy mountain peaks|a dense forest}|on a beach towel on a beach|on a {large|small} {red velvet|blue|pink|white|four post} bed|on a {white|black|blue|red|pink|purple|brown} couch|on a carpet|on a futon in a basement|in a mall|at a carnival|on stage in a theatre|in a barn on some hay|in the back of a delivery van}.

Lighting is {bright|soft|dim in the background|warm}.</code></pre>The only other trick I applied was training a low resolution 192x192. This tends to let HunyuanVideo learn the contours without keying into the specifics, assisting the model in generalization at render time. I also used dropout at 0.15, higher than I've gone before, given the limited dataset.<h2 id="disclaimer">Disclaimer</h2>blob

Files with SHA256: c3ef5c8b92388d0098999db3d0aab3f088b098390f8ef5ccd8f892cda289ac46