Deeply Optimized for Low/Mid-Range GPUs (12GB VRAM)

How to use:

• 1: Open ComfyUI and load the JSON

• 2: Load the starting image in the big green box

• 3: Type out your prompt int he big green box

• 4: Click on "Run" to generate a 30 second image to video

Warning: Your prompt has to be detailed. If it's something simple, it will just kind of rubberband on whatever simple inputs you describe, like "A man just sitting in the chair". The more details you add, the more it stitches together a seamless transition between the three independent shots to create a cohesive 30-second video in a single runtime pass. Then again, if all you wanted to do was make a simple generation, you wouldn't need a 30-second workflow.

The only thing the three shot separates do is dictate WHERE in the 30 seconds the actions take place. So the first set of quotations takes place within ten seconds; the second set of quotations take place within 20 seconds; the third set of quotations takes place after the 20 second mark.

Compromises had to be made to get this to run and generate in an acceptable time. It's possible to boost the image-to-video output for a sharper image, but you're looking at an average 21 minute render time at a step up in quality. Is it worth it? Depends on your workflow and if it's time sensitive.

This production-grade, crash-proof ComfyUI pipeline leverages Joey Gambino's advanced H3MultishotMemorySampler subgraph infrastructure. It has been systematically tuned to shatter the native 15-second tracking boundaries of the local MiniMax H3 architecture—successfully compiling up to 30 continuous seconds of 3-shot cinematic video with synced native audio tracks in under 15 minutes on a standard 12GB NVIDIA graphics card (such as an RTX 5070).

🛠️ Required Custom Node Packages

If any node blocks present a red warning threshold on your interface canvas, navigate to your ComfyUI Manager, execute Install Missing Custom Nodes, and restart your server environment. Alternatively, verify that the following core repository directories are fully initialized and updated:

1. comfyui-h3-multishot (By Joey Gambino)

2. Provides essential components: H3MultishotMemorySampler, H3ScriptSplit, H3ClipLoaderAny.

• ComfyUI-Spectrum-MiniMax-H3

• Provides essential components: SpectrumApplyMiniMaxH3 (Deploys advanced history parameters and signal stabilization to completely neutralize visual flickering).

• ComfyUI-FreeMemory

• Provides essential components: FreeMemoryImage (Acts as the system traffic cop to violently drop massive video models from memory prior to the video save cycle).

• comfyui-kjnodes

• Provides essential components: PathchSageAttentionKJ (Integrates highly optimized SageAttention mathematical libraries to keep GPU memory channels open).

📥 Required Model Inventory & Destination Paths

Ensure all specific neural weights listed below are manually stored within your local file tree. Modified nomenclature or inaccurate directory placement will result in model loading exceptions.

📂 Model Directory Map

markdown

📂 ComfyUI/
└── 📂 models/
    ├── 📂 vae/
    │   ├── 📄 minimax_h3_video_vae_fp16.safetensors
    │   └── 📄 minimax_h3_audio_vae_fp32.safetensors
    ├── 📂 diffusion_models/
    │   └── 📄 minimax_h3_fl2va_pruned_int8_convrot.safetensors
    ├── 📂 text_encoders/
    │   └── 📄 qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    └── 📂 loras/
        └── 📄 minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors

Use code with caution.

💾 Official Direct Asset Download Handles

• Video VAE (FP16): minimax_h3_video_vae_fp16.safetensors

• Audio VAE (FP32): minimax_h3_audio_vae_fp32.safetensors

• Diffusion Model Architecture: minimax_h3_fl2va_pruned_int8_convrot.safetensors

• Text Encoder Engine: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

• Turbo Model LoRA (8-Step Base): minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors

⚡ Mandatory Operational Environment Flags

To achieve absolute multi-shot stability and avoid unhandled Python environment abort failures during the long-form matrix sequence, you must explicitly configure your startup flags. Open your primary local execution script (e.g., run_nvidia_gpu.bat or initialization shell script) (or you can just open the ComfyUI desktop app and go to the Startup Args) and swap your launch command line argument array to match this configuration precisely:

bash

python main.py --disable-smart-memory --fp8_e4m3fn-text-enc --fp8_e4m3fn-unet 

Use code with caution.

Why these flags are mandatory:

• --disable-smart-memory: Mandates a hard PyTorch memory clean immediately upon raw clip finalization, bypassing background tensor leaks.

• --fp8_e4m3fn-text-enc: Compresses the massive 32B text encoder into lightweight 8-bit allocation blocks, locking it comfortably inside mid-range physical memory bounds.

📐 How to Achieve the 30-Second Long-Form Configuration

The workflow relies on a fine-tuned balance between your spatial layout constraints and frame processing intervals. Apply these precise configurations on the node face to duplicate the 14-minute execution baseline:

1. The Core Media Input: Drop your foundational tracking frame directly into the Load Image Here (Node 208) input bucket.

2. The Spatial Configuration: Inside ResolutionSelector (Node 115), anchor your values to 4:3 (Standard) with a megapixel evaluation slider locked cleanly at 0.4. This compact geometry drops pixel data overhead by more than 30% compared to heavy widescreen arrays, driving processing velocity forward.