Cobijada MiniMax-H3 Hybrid (Pruned, Uncensored)

A merged variant of MiniMax H3, a joint audio+video diffusion transformer (DiT) for ComfyUI, combining the strengths of the two source conditioning modes — first/last-keyframe (FL) and reference-driven (REF) — into a single checkpoint.

Why this merge exists

MiniMax H3 ships two distinct training regimes with identical architecture and tensor layout:

• FL (first/last-keyframe conditioning) — higher raw visual and audio output quality, but no support for reference-driven generation.
• REF (reference conditioning) — adds image/video/audio reference-driven generation, but at a real cost to output quality, even on non-reference tasks.

A tensor-by-tensor comparison of the two checkpoints shows the large majority of weights — attention projections, MLPs, norms, patch projections, rotary embeddings, and the token refiner — are effectively identical between them. The meaningful differences are concentrated almost entirely in the per-block adaln_proj weights: the AdaLN modulation projections that route conditioning signal into the residual stream at each transformer block.

That made a targeted, non-training weight-selection merge viable rather than a lossy compromise.

What this model is

This hybrid uses the FL checkpoint as the base for the entire network, with the adaln_proj weights for transformer blocks 25–49 (the later half of the network) taken from the REF checkpoint instead. Everything else — including the earlier-block adaln_proj weights, the final AdaLN projection, attention, MLPs, norms, and output heads — remains on FL throughout.

The intent is to keep REF's reference-conditioning pathway, which is expressed primarily through those later-block AdaLN weights, while retaining FL's higher-fidelity weights everywhere else in the network.

No additional training, fine-tuning, or gradient-based optimization was performed — this is a weight-selection merge at the tensor level, not a fine-tune. Both source checkpoints were the uncensored, pruned variants before merging.

Files

All three are pruned variants.

Usage (ComfyUI)

1. Download the format you want and place it in ComfyUI/models/diffusion_models/.
2. Load it with your normal MiniMax H3 checkpoint loader — architecture and tensor layout are unchanged from the source models, so no custom node is required to load the hybrid itself.
3. Reference-conditioned workflows (image/video/audio reference input) and standard first/last-keyframe workflows both work with this single checkpoint.

Intended use

• Text/image/video/audio-to-video generation where you want reference conditioning while keeping output quality closer to the FL checkpoint.
• A drop-in replacement for the REF checkpoint in reference-conditioned workflows, for users who found REF's raw output quality lacking.

This model is not expected to exceed the FL checkpoint's quality on non-reference-conditioned generation, since the large majority of its weights are shared with FL to begin with.

License

This merge, the quantized conversions, and this repository are released under Apache 2.0. The underlying MiniMax H3 weights remain subject to MiniMax's own original license terms — please review those separately before use.

Credits

Merged and quantized by Winnougan. Built on the officially released MiniMax H3 FL and REF checkpoints.