Sign In

Enzino Video2Prompt for MiniMax H3

Updated: Nov 19, 2026

toolminimaxwfpromptsvideoh3

Download

1 variant available

Archive Other

Enzino Video2Prompt v1.0.0.zip

355.97 KB

Verified:

Type
Workflows
Stats

116

Reviews
Published

Nov 19, 2026

Base Model

MiniMax H3

Hash
AutoV27D9AD803EE

The MiniMax H3 Community License Agreement’s grant excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America.

MiniMax H3

Enzino Video2Prompt - node 2.png

⭐ Unlock EnzinoAI Pro for just $25/month on EnzinoAI or Ko-fi ! Exclusive models, LoRAs, custom nodes & workflows. BF16, FP8, INT8 + new releases included.


📢 Join our Discord Community


Enzino Video2Prompt for MiniMax H3

A video in. An H3 prompt out.

Enzino Video2Prompt is a free ComfyUI node that watches any video clip and writes the prompt MiniMax H3 needs to recreate it, in MiniMax’s own official prompt format. Every shot change is timed to the millisecond. It runs 100% locally.

Drop a clip, press Run, and copy the prompt.

✨ What it does

  • Official H3 format. The prompt follows MiniMax’s own guide, with its three fields: integrated_multimodal_description, overall_soundscape and non_diegetic_music. H3 gets exactly the structure it was trained to read.

  • Every cut on the clock. The node finds the shot changes in your video and writes each one with its exact time: [Shot 2] At 00:02.000, the camera cuts to…. It also catches jump cuts where the framing stays the same, like a new outfit in a try-on video.

  • Long clips, still one prompt. H3 makes videos of 15 s at most. A longer clip is fitted into one 15 s prompt: every event is kept in order, and every cut time is scaled to the new length. You can also choose a window and keep only the part you want.

  • Camera language H3 understands. Push in, pull out, pan, truck, tilt, arc, tracking or static, with amplitude and speed: MiniMax’s own vocabulary.

  • Sound included. The node describes the likely ambience, action sounds and music from what it sees, because H3 generates audio. You can turn it off for a silent video.

  • Two modes:

    • Faithful (text-to-video): recreates the whole shot (style, framing, subjects, setting, action, camera, light).

    • Motion only (image-to-video): keeps just the action and the camera, so you can animate your own picture. The official H3 image-to-video line is added for you.

  • Notes field. Remake the shot your way, for example “replace the archers with two samurai, night, neon rain”. The timing and the camera moves of the original are kept.

  • Ready to wire. The outputs are prompt, first_frame (for image-to-video), duration (to set your H3 length) and frames.

🧠 Models: use what you already have

The model list shows what is already on your disk first (✓), then what can be downloaded (↓):

  • Qwen3-VL text encoders you already have (the ones used by Krea 2, Fibo 2, Qwen-Image 2.1…) work directly, with no download. They run through ComfyUI’s own engine.

  • Any Qwen vision model in models/VLM, models/LLM or your Hugging Face cache.

  • One-click download if you have nothing yet:

ModelSizeVRAMQwen3-VL-4B (recommended)8.9 GB~10 GBQwen3-VL-8B17.5 GB~7 GB in 4-bitQwen3-VL-2B (low VRAM)4.3 GB~5 GBQwen3.5-4B / 9B9.3 / 19.3 GB~10 / ~8 GB in 4-bit

The model is unloaded after each run, so H3 gets your whole GPU back.

📦 Install

  1. Copy Node/ComfyUI-Enzino-Video2Prompt into ComfyUI/custom_nodes/.

  2. Install the requirements with ComfyUI’s Python:
    python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-Enzino-Video2Prompt\requirements.txt

  3. Restart ComfyUI and load Workflow/Enzino Video2Prompt.json.

The full guide (settings, troubleshooting, tips) is in Guide/index.html.

  • format: H3 official

  • model: Qwen3-VL-4B, or the Qwen3-VL text encoder you already have

  • length: Detailed

  • frames: 16 (use 24–32 for many cuts or fast action)

  • detail: Fine (768 px), or Balanced (512 px) on 8 GB cards

🔓 Uncensored descriptions
Standard Qwen3-VL models may soften or skip some content. For a model that describes everything it sees as it is, use an uncensored (“heretic”) Qwen3-VL-4B text encoder: qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors. Drop it in models/text_encoders and pick it in the model list, where it shows up as “· text encoder”. No other setup needed.

💡 Tips

  • Wire duration to your H3 length, and first_frame to an image-to-video start image.

  • You can edit the prompt right in the node before you copy it.

  • One clip = one H3 video. For the best results, give it a clip of 15 s or less.

⚠️ Good to know

  • The model watches the video but can’t hear it: sound is inferred from the picture, and no dialogue is ever invented.

  • Only hard cuts are detected. Fades and dissolves count as a single shot.


Made by Enzino · enzinoai.com · Support my work on Ko-Fi

If you like it, a ❤️ and a review help a lot, and show me what you make with it!