Sign In
homemodelsimagesvideos3D Modelshubsarticlescomicschallengesupdatesshop

MiniMax H3 Remote Text Encoder - ComfyUI Node (Low VRAM) - v1.6.0 Showcase

jeffamine's Avatar

jeffamine

Loading Images

Short description

Run MiniMax H3's 32B Qwen3-VL conditioner from ComfyUI through free Hugging Face ZeroGPU, with an optional automatic private Kaggle GPU fallback and local conditioning cache.

Full description

This unofficial custom node replaces the local MiniMax H3 text-encoder loader with a cloud-backed CLIP-compatible loader. It supports text, H3 image-to-video keyframes, and H3 reference-to-video inputs without keeping the 32B Qwen3-VL encoder in local VRAM.

The default backend is the public multimodalart/qwen3vl-conditioner Hugging Face Space. Users can run anonymously or paste a Hugging Face token directly into the loader's masked setup panel. When Hugging Face is out of quota or temporarily unavailable, the node can automatically use the user's own free Kaggle GPU quota.

Kaggle setup is account-independent: paste a Kaggle API token into the second masked field and click Save & test tokens. The fields are cleared after saving and are never serialized into the workflow. The account name is detected automatically. Each request creates a uniquely named private notebook, downloads the conditioning, and attempts to delete the notebook after completion or failure.

Successful conditioning is cached locally, so identical prompts and references do not consume cloud quota again. Kaggle uses the official pinned MiniMax H3 NVFP4 AWQ encoder and may differ numerically from the BF16 Hugging Face backend.

Tested end to end with text, image references, automatic Hugging Face-to-Kaggle failover, and a combined video-plus-audio reference request. The multimodal test returned finite [1, 1068, 5120] conditioning with valid text/media tags and left no temporary Kaggle notebook behind.

Install

1. Extract the included ComfyUI-H3-Remote-Encoder folder into ComfyUI/custom_nodes/.

2. Restart ComfyUI and add MiniMax H3 Remote Encoder Loader.

3. Click Install/check components inside the node.

4. Paste tokens into the masked fields and click Save & test tokens.

5. Connect the loader to the native MiniMax H3 conditioning node.

See README.md and README_RU.md in the archive for setup, privacy details, limitations, and a test workflow.

Privacy and external services

Prompts and selected references are uploaded to the chosen cloud provider. Kaggle jobs are private and cleanup is attempted automatically. A force-killed local process can leave an orphaned private notebook that the user must delete manually. Free-service quotas, queues, availability, and terms can change.

No model weights, credentials, author account, or machine-specific paths are included. Source code is MIT licensed. MiniMax H3, ComfyUI, Hugging Face, Kaggle, and hosted assets retain their own licenses and terms. This project is not affiliated with those services.

Comments