MiniMax H3 16 GB Low-VRAM ComfyUI Workflow (text-to-video, Turbo 6 steps)
=========================================================================

Article: https://freakyjolly.com/minimax-h3-out-of-memory-comfyui-16gb-fix/

What's inside:
  - MiniMax_H3_T2V_16GB_LowVRAM.json: ComfyUI workflow (UI format). Drag it onto the canvas.
  - MiniMax_H3_T2V_16GB_LowVRAM_api.json: Same graph in API format.

Required models (download each file and put it in the folder shown):
  - minimax_h3_fl2va_pruned_int8_convrot.safetensors
      folder: ComfyUI/models/diffusion_models/
      download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
  - minimax_h3_turbo_4step_ema_ckpt500.safetensors
      folder: ComfyUI/models/loras/
      download: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/resolve/main/minimax_h3_turbo_4step_ema_ckpt500.safetensors
  - qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
      folder: ComfyUI/models/text_encoders/
      download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
  - minimax_h3_video_vae_int8_convrot.safetensors
      folder: ComfyUI/models/vae/
      download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors
  - minimax_h3_audio_vae_fp32.safetensors
      folder: ComfyUI/models/vae/
      download: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors

Custom nodes:
  - ComfyUI-MiniMax-H3-Turbo (MiniMaxH3TurboLoRA, MiniMaxH3TurboSampler): https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo  (ComfyUI Manager > Install Missing Custom Nodes, or git clone into ComfyUI/custom_nodes/)
  - ComfyUI-KJNodes (VRAM_Debug, MiniMaxLowVRAMAttention, MiniMaxChunkFeedForward, PathchSageAttentionKJ, ModelPreviewOverrideKJ): https://github.com/kijai/ComfyUI-KJNodes  (ComfyUI Manager > Install Missing Custom Nodes, or git clone into ComfyUI/custom_nodes/)
  - ComfyUI-SuperNodes (SetReserveVRAM): https://github.com/sonnybox/ComfyUI-SuperNodes  (ComfyUI Manager > Install Missing Custom Nodes, or git clone into ComfyUI/custom_nodes/)
  - ComfyUI-Easy-Use (easy cleanGpuUsed, easy clearCacheAll): https://github.com/yolain/ComfyUI-Easy-Use  (ComfyUI Manager > Install Missing Custom Nodes, or git clone into ComfyUI/custom_nodes/)
  - ComfyUI-VideoHelperSuite (VHS_VideoCombine): https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite  (ComfyUI Manager > Install Missing Custom Nodes, or git clone into ComfyUI/custom_nodes/)

How to use:
  1. Install the custom nodes above and restart ComfyUI. Recommended launch flags from the article: --reserve-vram and --cache-none.
  2. Download the model files into the folders listed.
  3. Drag MiniMax_H3_T2V_16GB_LowVRAM.json onto the canvas.
  4. Edit the prompt, keep 0.3 MP and about 5 s for a first run, and queue. Output: ComfyUI/output/video/MiniMax_H3_t2v_16gb*.

Notes:
  - This is my text-to-video Turbo graph with the fixes from the article applied: UNETLoader with the pruned int8_convrot model, SetReserveVRAM 3.0 after CLIPLoader, KJNodes VRAM_Debug after conditioning (empty_cache + gc_collect), MiniMaxLowVRAMAttention + MiniMaxChunkFeedForward (seq threshold 4096) before sampling, 0.3 MP and 5 s.
  - This exact combined graph has not been benchmarked; the fixes are the ones described stage by stage in the article.
  - If text encoding still stalls, swap CLIPLoader for the 4B int4 encoder + ClipProj used in the text-to-video workflow.
  - Sage Attention is "auto"; set it to "disabled" if SageAttention is not installed.

Model weights are not included. Each model keeps its own licence; check the model page before commercial use.
Workflow shared by FreakyJolly.com.
