MiniMax-H3 Ref2VA Ultra-Detail Workflow (Merged Model & Dual-Sampler)

This is currently what I consider the best workflow for MiniMax-H3 multi-reference video generation! Designed specifically to overcome high-motion artifacting while significantly boosting video clarity, this workflow features two key innovations:

• Dual-Stage Resampling via ExtendIntermediateSigmas: Fixes video artifacts and injects rich details under high-motion conditions by adding linear noise interpolation to the final 2 steps.

• Custom Merged Diffusion Model: Powered by a hybrid checkpoint merging fl2va (for superior image quality) and ref2va (for temporal stability), giving you pristine visual fidelity while fully retaining reference video editing and audio capabilities.

☁️ Try It Online (Free Test Drive)

Test this workflow directly in your browser on RunningHub before downloading:

👉 Run on RunningHub

⚙️ Recommended Model Configuration

1. Diffusion Model (Base)

• Model: minimax_h3_ref2va_XUELUO_int8_convrot (High-fidelity merged Ref model)

• Download: Civitai Model Page (Currently uploading; link will be live ~1 hour post-release)

• Download2:WarmBloodAban/Minimax_h3_ref2va_XUELUO_int8_convrot · Hugging Face

2. Text Encoder, VAE & Acceleration LoRA

• Text Encoder: qwen3vl_32b_minimax_h3_int8_convrot

• 4-Step Acceleration LoRA: minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16

• VAE: minimax_h3_video_vae_int8_convrot

• Download Sources:

  • HuggingFace: Comfy-Org/MiniMax-H3

  • HuggingFace: Kijai/MiniMax-H3-Playground

💡 Generation & Tuning Tips

• Standard Speed Run (Recommended): Set steps to 4. The workflow will automatically append 2 additional interpolation steps at the end to refine details and smooth out noise.

• Ultra-High Quality Setup:

  • Increase initial steps to 6–8 (or bypass the 4-step LoRA completely and run at 20 steps).

  • Scale up the scale_by parameter in the Upscale Latent By node from 1.2 to a higher value (Note: requires significantly higher VRAM / GPU compute).

If you find this workflow helpful, please consider giving it a 👍 and sharing your generated runs in the gallery!