CoolFace
Modelpublic

Abiray/MiniMax-H3-Pruned-Ref-Delta-Fused-GGUF

sourceHugging Faceotherupdated 8d agoView on Hugging Face
6likes1.2kdownloads
Model Card

MiniMax-H3 Pruned Ref-Delta Fused (GGUF)

GGUF quantizations for xmarre's MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-ComfyUI.

This hybrid model fuses the crisp visual quality, motion dynamics, and clean audio of FL2VA with the persistent multi-image character conditioning of Ref2VA using a rank-1024 delta weight merge ($\Delta W_{r=1024}$) on the pruned 20.1B parameter MiniMax-H3 architecture.

With these GGUFs, you can now run reference-conditioned MiniMax-H3 locally on consumer GPUs (16 GB – 24 GB VRAM) with minimal quality loss.

File Overview

FilenameSizeRecommended GPUDescription
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q8_0.gguf`21.6 GB24 GB+ / Multi-GPUNear-indistinguishable from BF16 base
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q6_K.gguf`16.7 GB24 GB VRAMHigh-fidelity sweet spot for critical renders
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q5_K_M.gguf`14.1 GB16 GB – 24 GB VRAMExcellent balance of motion consistency and size
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q5_K_S.gguf`14.1 GB16 GB VRAMSlightly smaller footprint with high precision
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q4_K_M.gguf`11.6 GB16 GB VRAMMost Recommended. Fast, low VRAM, sharp details
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q4_K_S.gguf`11.6 GB16 GB VRAMLightweight 4-bit quantization
`MiniMax-H3-Pruned-Ref-Delta-Fused-Q3_K_M.gguf`8.9 GB12 GB – 16 GB VRAMMaximum VRAM savings for constrained setups

How to Use in ComfyUI

1. Install Node

Ensure you have the latest ComfyUI-GGUF custom node installed via ComfyUI Manager.

2. Place Files

Place the downloaded .gguf file into: ComfyUI/models/diffusion_models/ (or `ComfyUI/models/unet/`)

Recommended Generation Settings

With Turbo Acceleration (Fast Inference)

  • —LoRA: minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_resized_avg_rank_20_bf16.safetensors
  • —LoRA Strength: 1.0 (or 0.85 if fine ripple noise occurs)
  • —Steps: 8 steps
  • —CFG: 1.0 (Do not exceed 1.2 with distilled turbo LoRAs)
  • —Sampler: Euler
  • —Scheduler: Beta or Simple
  • —Resolution: 1280x720 (16:9) or 1280x544 (2.39:1)
  • —Frame Count: 49 or 81 frames ($4k + 1$ rule)

Credits & Acknowledgements

  • —Base Architecture: MiniMax AI
  • —Ref-Delta Low-Rank Merge: xmarre
  • —ComfyUI GGUF Integration: city96
  • —GGUF Conversion & Quantization: Abiray