CivArchive
    FisherKing-MiniMax-H3-[4Steps]-T2V-ReferenceWorkflow-v1.0 [Low VRAM] - v1.0
    NSFW
    Preview 139235267

    Workflow Goal:

    Provide a clean reference implementation for MiniMax H3 based text-to-video generation.

    This workflow focuses on simplicity, reproducibility, and education rather than including every available feature or optimizations. It reflects a tested baseline for producing high-quality videos from a prompt. Treat it as a starting point and adapt the prompt and LoRA stack for your preferred artistic direction.

    IMPORTANT NOTE:

    Validated using an NVIDIA RTX 2080 (8 GB VRAM) with 64 GB System RAM.

    Low VRAM does not mean low hardware requirements.

    MiniMax H3 Text-to-Video is computationally intensive. Even when GPU VRAM is managed efficiently, higher resolutions and longer videos can require substantial system RAM for memory paging and intermediate data.

    ## Model

    - MiniMax H3 INT8 quantized diffusion model

    https://civarchive.com/models/2830065/minimax-h3-int8int4-convrot

    - Qwen3-VL 32B NF4/AWQ text encoder

    https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

    - H3 Video VAE FP16

    https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors

    - H3 Audio VAE FP32

    https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors

    - MiniMax H3 Turbo LoRA

    https://civarchive.com/api/download/models/3206543?fileId=3088145

    ## Tested Configuration

    - GPU: NVIDIA RTX 2080 8 GB

    - System RAM: 64 GB

    - Resolution: 768 × 512

    - Duration: 5 seconds

    - CFG: 1.0

    - Sampler: res_multistep

    - Scheduler: simple

    ## Steps

    4 steps + Turbo LoRA

    - Primary configuration.

    - Good visual quality, motion, effects and audio.

    - Very fast relative to the base model.

    - Slight slow-motion feel may occur at the beginning of some clips.

    ## Optimization

    - No hardware-specific attention optimizations used.

    - No SageAttention.

    - No other VRAM/speed optimization nodes or launch parameters required.

    - Reference workflow intentionally uses the standard H3/ComfyUI execution path.

    ## Credits

    - MiniMax H3: MiniMax

    - H3 ComfyUI implementation / model work: Kijai

    - Turbo LoRA: LightX2V

    - INT8 quantization: see model source linked above

    - Workflow reference: MiniMax H3 community/reference workflow

    Description

    FAQ

    Workflows
    MiniMax H3

    Details

    Downloads
    198
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/9/2026
    Updated
    8/13/2026
    Deleted
    -

    Files

    fisherkingMinimaxH34steps_v10.json

    Mirrors