Community-trained anti-burn LoRA fixes face degradation in chained H3 video generations
TypeItRight · reddit · 2026-10-16
- When chaining H3 clips into longer videos, each continuation copies the last clip's flaws and adds more — by clips 5-10, faces turn crunchy, etched, and over-sharpened.
- Fix: trained with a community compute grant, this LoRA loads only on continuation clips (skip the first one) to keep them close to clip 1's look.
- Results (37s chains, face texture change clip 1→10): lightx2v 8-step +44%→0%; HyperFlow +25%→−1%; an untrained model +61%→+4%; Larry Turbo +261%→−5% (can soften heavily burned output).
- Usage: LoraLoaderModelOnly on continuation clips, start at 0.5; go higher for harder-burning turbos, lower if faces soften.
- Weights, training code, recorder node, and scoring script are open-sourced on HF and GitHub; the author is soliciting feedback on other turbos.
More from Multimodal
- AI-Rendered Gotham After Dark: Retro Fabric-Print Bat Signals and Walls of Joker Faces — egeberkina · 2026-10-16
- Turn charts, logos and photos into animated videos in 5 minutes with Claude Code + Remotion — 6 copy-paste prompts — TawohAwa · 2026-10-16
- Kling 4.0's Omni Reference supports up to 15 reference assets in one generation — SarahAnnabels · 2026-10-16
- Kling 4.0 holds character and environment coherent in large-scale two-person climbing shots — SarahAnnabels · 2026-10-16
- Kling 4.0 keeps character stable across a massive desert scene with endless movement — SarahAnnabels · 2026-10-16
- Seedance 2.5 workflow: character reference sheets + structured prompt yields a 30s single-take fake tourist video — techhalla · 2026-10-16