Downloads · 30 days
18.9K
45% of all-time downloads
sandpies/Minimax-H3-fl2va-ref2va-hybrid-unpruned-int8
Minimax-H3-fl2va-ref2va-hybrid-unpruned-int8 is a machine learning model from sandpies. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for minimax-h3. The card lists the license as other.
Community weight-selection merge of the unpruned MiniMax H3 int8convrot checkpoints from Comfy-Org/MiniMax-H3.
Downloads · 30 days
18.9K
45% of all-time downloads
All-time downloads
41.6K
Public
Repo size
68.1 GB
Likes
10
Public
Click a slice to open those files.
.safetensors68.1 GB · 100%
From the Hugging Face model README
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).
No training. For each tensor:
minimax_h3_fl2va_int8_convrot.safetensors (unpruned)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| File | Ref2VA AdaLN blocks | Tradeoff |
|---|---|---|
minimax_h3_hybrid_fl2va_ref2va_b25-49_unpruned_int8.safetensors | 25–49 (25 blocks) | More reference pathway / slightly less FL2VA-pure |
minimax_h3_hybrid_fl2va_ref2va_b30-49_unpruned_int8.safetensors | 30–49 (20 blocks) | Closer to FL2VA quality / slightly weaker ref AdaLN |
Final-layer AdaLN and output heads remain FL2VA in both.
| smhfacct hybrids | this repo | |
|---|---|---|
| Source family | pruned INT8 (adaln_t_table) | unpruned INT8 (time_embedder) |
| Approx size | ~21 GB | ~32 GB |
| Merge recipe | FL2VA + late-block Ref2VA AdaLN | same recipe, different body |
Do not mix pruned and unpruned files in one load.
ComfyUI/models/diffusion_models/UNETLoaderMiniMaxH3ReferenceToVideo, HandTie, etc.)Optional: the same merge can be done at runtime with scottmudge/ComfyUI_MinimaxH3HybridLoader (block_range_adaln).
Reference-conditioned audio+video generation where you want FL2VA-like fidelity with Ref2VA’s late-block AdaLN reference routing, on the unpruned INT8 stack.