This is a high-performance, multi-stage video generation workflow designed specifically for the Minimax H3 Eros Max architecture. By utilizing a dual-pass sampling strategy combined with advanced frame interpolation and 3D latent upscaling, this workflow bridges the gap between standard generative video and cinematic-quality output.

The pipeline is engineered for professional creators, featuring an organized node structure, real-time visual feedback via live preview, and optimized LoRA management to streamline the creative process.

🚀 Key Features

Dual-Pass Sampling Engine: A sophisticated two-stage generation process that ensures temporal consistency in the first pass and high-frequency detail refinement in the second.

Advanced Upscaling & Interpolation: Integrates MinimaxH3LatentUpscaler3D for spatial resolution enhancement and FILM (Frame Interpolation for Large Motion) for buttery-smooth frame rates.

Hybrid Input Modes: Full flexibility with support for Text-to-Video, First-Frame conditioning, and Last-Frame conditioning for seamless video looping or extended generation.

Smart LoRA Management: Includes a specialized folder-based LoRA loader, allowing you to select models from specific directories—eliminating the need to scroll through massive lists in ComfyUI.

Real-Time Monitoring: Features an integrated live video preview on the main workspace, allowing you to monitor generation progress and motion dynamics without switching windows.

Optimized Speed: Built with speedups in mind, utilizing Triton-based attention for faster sampling cycles.

🛠 Technical Requirements

*Required Custom Nodes*

To ensure the workflow runs correctly, please install the following via ComfyUI Manager:

PlagueKind-Nodes
KJNodes
SolAttn_triton / sol-attn
ClipProj
MinimaxH3LatentUpscaler3D
LBH-123-AI/Minimax_h3_latent_Upscaler

*Required Models*:

Component Model Name
Diffusion Model H3 Eros Max
CLIP Model qwen3-vl-4b-heretic_nvfp4
VAE Minimax Audio & Video VAE
Upscaler minimax_h3_latent_upscaler_3d_bf16.safetensors
Interpolation film_net_fp16.safetensors

⚙️ Workflow Architecture

Conditioning Stage: User inputs text prompts or image references (First/Last frame). The LoRA loader pulls specific styles from your designated folders.

Base Generation (Pass 1): A low-resolution sampling pass establishes the motion trajectory and temporal structure using optimized SolAttn kernels.

Refinement & Upscale (Pass 2): The latent data is passed through a 3D Latent Upscaler, followed by a second high-detail sampling pass to inject texture and clarity.

Temporal Smoothing: The FILM interpolation node processes the upscaled frames to increase the final FPS, resulting in fluid, professional motion.

💡 Pro-Tips & Usage Notes

Resolution vs. Speed: While this workflow supports high-resolution outputs (e.g., 1.0 Megapixel), please note that the second pass step time increases exponentially at higher resolutions. For rapid prototyping, it is recommended to generate at lower resolutions and use the upscaler for final renders.

LoRA Organization: Use the included LoRA loader to point to specific sub-folders in your models directory to keep your workspace clean and your workflow fast.

Hardware Recommendation: Due to the 3D Latent Upscaling and dual-pass sampling, a high-VRAM GPU (24GB+ recommended) is highly suggested for optimal performance at high resolutions.

If you have any questions beyond this description, let me know!