berryber09/MiniMax-H3-ref2va-fl2va-hybrid-w4a8

MiniMax-H3 · fl2va × ref2va Hybrid (w4a8, b20-49)

A 4-bit (w4a8_mixed) block-splice hybrid of MiniMax-H3 that combines the first/last-frame keyframe model (fl2va) with the reference-conditioning model (ref2va) in a single checkpoint — so one gen can be driven by a first frame and use reference images at the same time.

~12.5 GB on disk — 4-bit (w4a8_mixed).

Why

• fl2va — first/last-frame keyframe conditioning, higher visual/audio quality. No reference slots.
• ref2va — image/video/audio reference conditioning (subject/scene/style). Lower base quality.

Neither alone can do "open on this frame, and have this person walk in later." The hybrid can: the first frame stays a hard temporal anchor while the reference supplies a subject's identity that the model can introduce anywhere in the clip.

Recipe (b20-49)

Whole-block splice of the two official-derived w4a8 checkpoints:

Part	Source
Input/output heads + all non-block tensors	fl2va
Transformer blocks 0–19	fl2va
Transformer blocks 20–49	ref2va

b20-49 leans toward reference adherence (30 of 50 blocks from ref2va) at a small cost to base quality — a good pick when the reference is mandatory. The quality↔reference spectrum, most-quality → most-reference: b30-49 → b25-49 → b20-49 → b15-49.

Per-layer quant metadata is merged correctly: fl2va's _quantization_metadata is kept only for the fl2va-sourced layers; the ref2va blocks carry their own comfy_quant tensors. Verified to load in ComfyUI (Detected mixed precision quantization).

Usage (ComfyUI)

Place in ComfyUI/models/diffusion_models/. It routes as a ref2va model, so both paths are active:

• first_frame / last_frame reference → wired as a keyframe (MiniMaxH3AddKeyframes) — hard-anchors that frame.
• subject / scene / style reference → wired into MiniMaxH3ReferenceToVideo — identity/style conditioning, not frame-locked.

Both can be supplied in one generation.

Credit

• Base model: MiniMaxAI/MiniMax-H3
• Hybrid block-splice idea: smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models (int8 / bf16-sized). This repo adds the w4a8 variant, which that repo does not provide.

Inherits the MiniMax-H3 license.