CoolFace
Modelpublic

sandpies/Minimax-H3-fl2va-ref2va-hybrid-unpruned-int8

sourceHugging Faceotherupdated 11d agoView on Hugging Face
5likes2.3kdownloads
Model Card

MiniMax H3 Hybrid (unpruned INT8) — FL2VA base + Ref2VA AdaLN

Community weight-selection merge of the unpruned MiniMax H3 int8_convrot checkpoints from Comfy-Org/MiniMax-H3.

Same hybrid idea as smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models, but built from the full / unpruned INT8 pair (time_embedder) instead of the pruned pair (adaln_t_table).

What’s in these files

No training. For each tensor:

  • —Base: minimax_h3_fl2va_int8_convrot.safetensors (unpruned)
  • —Overlay: blocks.{N}.adaln_proj.linear.{weight,bias} (and matching .comfy_quant / _scale siblings) taken from minimax_h3_ref2va_int8_convrot.safetensors (unpruned) for a late block range
  • —Everything else stays FL2VA
FileRef2VA AdaLN blocksTradeoff
minimax_h3_hybrid_fl2va_ref2va_b25-49_unpruned_int8.safetensors25–49 (25 blocks)More reference pathway / slightly less FL2VA-pure
minimax_h3_hybrid_fl2va_ref2va_b30-49_unpruned_int8.safetensors30–49 (20 blocks)Closer to FL2VA quality / slightly weaker ref AdaLN

Final-layer AdaLN and output heads remain FL2VA in both.

Pruned vs unpruned

smhfacct hybridsthis repo
Source familypruned INT8 (adaln_t_table)unpruned INT8 (time_embedder)
Approx size~21 GB~32 GB
Merge recipeFL2VA + late-block Ref2VA AdaLNsame recipe, different body

Do not mix pruned and unpruned files in one load.

Use in ComfyUI

  1. 1.Put a file in ComfyUI/models/diffusion_models/
  2. 2.Load with stock Load Diffusion Model / UNETLoader
  3. 3.Treat the model as Ref2VA for reference packing (MiniMaxH3ReferenceToVideo, HandTie, etc.)

Optional: the same merge can be done at runtime with scottmudge/ComfyUI_MinimaxH3HybridLoader (block_range_adaln).

Intended use

Reference-conditioned audio+video generation where you want FL2VA-like fidelity with Ref2VA’s late-block AdaLN reference routing, on the unpruned INT8 stack.

Limitations

  • —Experimental / subjective block ranges; try b30–49 first if you care more about raw quality, b25–49 if reference adherence feels weak.
  • —Inherits MiniMax H3 licensing and usage terms from the source checkpoints. This merge grants no extra rights.

Acknowledgements

  • —MiniMax for MiniMax H3
  • —Comfy-Org for the published INT8 convrot weights
  • —smhfacct for documenting / shipping the pruned hybrid approach this mirrors