H3 Anti-Burn LoRA keeps faces from burning in chained MiniMax H3 clips
H3 Anti-Burn LoRA is an open LoRA for MiniMax H3 that you load only on continuation clips, so a long chained video keeps the look of its first clip instead of getting crunchier faces with every segment. On a turbo model it was never trained on, its author measured face burn after ten clips drop from +61% to +4%.
What it does
When you build a long H3 video by chaining clips, each clip continues from the end of the previous one, copies its small flaws and adds a few more, so after a few clips skin and lines look etched and the picture gets harsher. The LoRA was trained to make a continuation look like a fresh generation instead of amplifying the flaws in its context.
You load it with LoraLoaderModelOnly on the model that makes clips 2, 3, 4 and so on, never on clip 1, and start at a strength of 0.5.
Results and limits
The author ran 10-clip chains of about 37 seconds, four scenes per setting. Face burn went from +25% to −1% on HyperFlow 8-step at 0.5, from +44% to 0% on lightx2v 8-step at 0.5, and from +61% to +4% at 0.6 on a model it never saw in training. Larry Turbo v4 burns far harder and needed more strength; too much strength gives soft, waxy faces, and fine detail drops a little.
Worth knowing
It's experimental: tested on speaking people at medium to close distance and on continuations from masked context, and it prevents burn but doesn't repair a video that's already burned. Merging it into an 8-bit checkpoint weakens it, so try a few strengths on a 4-5 clip chain. The training code is open, and the 310 MB file is a Model Derivative of MiniMax H3 under the MiniMax H3 Community License: Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use.
Try it yourself
In your H3 continuation workflow, add LoraLoaderModelOnly with h3_antiburn_universal_v1.safetensors at 0.5 on the model for clips 2 and later only, run a [5]-clip chain of [your dialogue scene], and raise the strength if faces still get crunchier from clip to clip.