Download
1 variant available
Config Other
FisherKing-MiniMax-H3-[4Steps]-T2V-ReferenceWorkflow-v1.0.json
22.19 KB
Verified: 2 months ago
11.8K
574
168
Generation, training and LoRA distribution on Civitai are covered by Civitai’s own license agreement with MiniMax. If you download these weights and run them yourself, your use is instead governed by the MiniMax H3 Community License Agreement, whose grant excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America.
MiniMax H3
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://civitai.red/models/2830065/minimax-h3-int8int4-convrot
- Qwen3-VL 32B NF4/AWQ text encoder
- 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://civitai.red/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