CleanTie & HandTie
V2Sing Hybrid — MiniMax-H3 Hybrid Checkpoints for ComfyUI
2 min read
A new build of hybrid weights and LoRA adapters for video and audio generation based on MiniMax-H3, designed for work in ComfyUI. The model combines the FL2VA and Singularity V2 architectures, allowing it to maintain stable reference and color consistency without degrading the audio branch. Options are available with a pre-integrated 8-step turbo mode and calibrated anatomy (including SFW and NSFW versions). This set is suitable for video artists and enthusiasts working with complex video generation pipelines that support audio.
Key features
- Hybrid architecture: Combines FL2VA and Singularity V2 with closed-form AdaLN alignment (Adaptive Layer Normalization — a neural network layer adaptation method).
- 8-step turbo: Pre-installed turbo versions and separate LoRA adapters for fast generation in just 8 sampling steps.
- Audio safety: A refined backend (PRISM-v2) protects the model's audio branch from degradation during compression and quantization.
- Clean and HandTie options: Separation into safe SFW versions (Clean) and builds with improved anatomy (HandTie).
What you can do
- Video generation: Create detailed video clips from images and references in ComfyUI.
- Pipeline acceleration: Use ready-made `_turbo_` checkpoints for fast rendering in 8 steps without extra LoRA adapters.
- Fine-tuning: Combine base weights with standalone LoRA modules to control style and anatomy.
- VRAM savings: Run INT8 ConvRot quantized versions on compatible hardware.
Example
Recommended settings for ComfyUI: 8 steps, CFG 1.0, lcm or euler sampler, along with shift parameters shift_video = 12.0 and shift_audio = 2.5.
Read more
DetailsBenchmarks: uses weights that preserve >=99% Frobenius energy.
Hardware requirements: available in BF16 format (approx. 37.46 GiB) and INT8 ConvRot (approx. 20.59 GiB).
Training details: closed-form AdaLN alignment using float64 least-squares.
License: MiniMax-H3 Community License Agreement.














Comments