Cauthyyy/minimax-h3-ref2va-community-comparison

MiniMax H3 — Reference Performance

Experiment ID: refcommunity20260917

An independent comparison of the original Ref2VA model at 50 actual forward evaluations and five community 8-evaluation configurations, across ten mixed-order public reference-generation/editing tasks.

Each task displays all input images/videos/audio, the exact prompt, six generated audiovisual results, downloadable settings, and qualified observation notes. Explicit initial video and audio noise is identical within each task. No result borrows another model's soundtrack.

Rkss uses the publisher-recommended FL2VA–Ref2VA hybrid checkpoint, not the original Ref2VA base. This is a comparison of complete configurations, not a pure same-base LoRA ablation or an official leaderboard. The webpage documents scheduler and runtime distinctions.

Native-format adapters and the hybrid base are converted with the official gate/value row permutation. An earlier incorrect conversion was detected during review; those outputs were invalidated and rerun with the same fixed inputs. Published results use the corrected conversion and carry component-level numerical audits. These checks do not establish end-to-end pixel parity with a separate ComfyUI installation.

The reference inputs are public examples from LightX2V, MiniMax and fal. Source links are retained per task and per asset. Rights to third-party models and reference media remain with their respective owners; no ownership of those assets is claimed. Model licenses and applicable use policies continue to apply.

Static HTML/CSS/JavaScript; no analytics, external UI dependencies or credentials. For local preview, run python -m http.server 8765 in this directory.