Low Poly LoRA (rank 32, 5000 steps) + model card, sample

c734559 verified 10 days ago

2.05 kB

license: other
license_name: minimax-h3-community-license
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
base_model: MiniMaxAI/MiniMax-H3
base_model_relation: adapter
pipeline_tag: image-text-to-video
tags:
  - lora
  - video
  - text-to-video
  - low-poly
  - retro
  - 3d
  - 90s
  - style
  - minimax-h3

Low Poly — retro low-poly 3D for MiniMax H3

A style LoRA (rank 32) that renders MiniMax H3 footage as late-90s console-era 3D — faceted low-poly geometry, flat-shaded surfaces, chunky texture pixels, and a bright, saturated palette. H3 trains audio jointly, so dialogue and ambience come through with the picture.

This repo hosts the weights for the fal endpoint, which is the supported way to use this style — the endpoint applies the tuned trigger phrasing and settings server-side:

minimax/h3-max/lora-gallery/low-poly on fal.ai

Raw single take at 768P / 4:3, picture and voice generated together in one pass — no cherry-picking.

Files

file	what
`low-poly.safetensors`	the LoRA, rank 32, 5,000 steps
`samples/used-car-ad.mp4`	the example clip above

Training

Trained with fal's MiniMax H3 t2v trainer (rank 32, 5,000 steps, 4:3, joint video+audio objective) on a small hand-curated set of 52 three-second clips of native low-poly-era console gameplay captured losslessly at 4:3, each clip human-reviewed for style before inclusion. The 5,000-step checkpoint was chosen over its 3k/8k/10k siblings by building a complete six-scene practical showcase and a 15-second trailer with it and comparing real-world results.

License

Adapter weights for MiniMaxAI/MiniMax-H3, distributed under the MiniMax H3 Community License.