Home Creators Posts Import Register Favorites Logout

Content

Video : https://youtu.be/JMDPignEbP0

Hey everyone!

New video is up https://youtu.be/JMDPignEbP0 , and this one is the full Tencent Hy-Nomas setup in ComfyUI.

Hy-Nomas is a world model, so instead of generating a clip and forgetting everything, it keeps the environment in memory. Same buildings, same banners, same lighting, across one continuous shot. In the video I run a cyberpunk walk-through where an orange banner stays put through a 20-second camera move, and a 40-second first-person IG influencer video going from a shopping mall to a park, stitched from four segments in one pipeline.

Honest finding: the Drive Interactive mode, where you steer the world with WASD keys in real time, is laggy on a 24 GB card and eats memory fast. Fun to try once, not usable for real work. The pre-recorded camera control is the one that actually works.

For Patreon Supporter, 2 workflows I like personally and some examples JSON.

https://www.patreon.com/aifuturetech/posts/hy-nomas-worlds-174376828

Thanks for watching. See you in the next one.

— Benji

Resources:

- Hy-Nomas JSON prompt skill: https://github.com/benjiyaya/hy-nomas-prompts

- Hy-Nomas open weights: https://huggingface.co/tencent/Hy-Nomas

- Project page: https://hy-nomas.hy.tencent.com/

- ComfyUI-Hy-Nomas custom node: https://github.com/PedroMarinhoDev/ComfyUI-Hy-Nomas/

Workflows: https://github.com/PedroMarinhoDev/ComfyUI-Hy-Nomas/tree/main/workflows

- ComfyUI-packaged model files: https://huggingface.co/PedroMarinhoDev/Hy-Nomas-ComfyUI/tree/main

- Qwen 2.5 VL 7B text encoder: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI

Files

Hy-Nomas Video World Model in ComfyUI - Running Long Video Easily!

This video walks through ComfyUI Hy-Nomas, the custom node pack that runs Tencent's Hy-Nomas open-weights video world model natively in ComfyUI. It covers the full setup (the three model files: distilled diffusion model, fp8 VAE, and Qwen 2.5 VL 7B text encoder), the bundled example workflows, the Hy-Nomas Drive Interactive keyboard mode, start-frame workflows, the timestamp and event-driven JSON prompt format, extending worlds across multiple segments for long videos, overriding segment prompts, and pre-recorded WASD camera-control trajectories. For ComfyUI users and AI video creators who already run models like Wan 2.1, Wan 2.2, Qwen Image, or MiniMax H3, and want to try a true world model locally. If you have a 24 GB VRAM GPU (RTX 4090 class) and you are tired of stitching loops to fake long coherent videos, this is the workflow for you. World models remember their environment. Instead of forgetting the scene every five or ten seconds like standard text-to-video models, Hy-Nomas keeps history frames, so objects, banners, and buildings stay coherent across a single one-shot take of 20, 40 seconds or longer. It runs in 24 GB of VRAM instead of the 80 to 100 GB that most real-time world models need, which makes local world-model video generation actually possible right now. Resources: Post: https://www.patreon.com/aifuturetech/posts/hy-nomas-video-174376205?utm_medium=video&utm_source=youtube&utm_campaign=202611 - Diffusion model (distilled safetensors): https://huggingface.co/PedroMarinhoDev/Hy-Nomas-ComfyUI/tree/main Goes into ComfyUI/models/diffusion_models - VAE (fp8 safetensors): same repo, goes into ComfyUI/models/vae - Text encoder: Qwen 2.5 VL 7B, from ComfyUI official repo https://huggingface.co/Comfy-Org/Qwen-Image/tree/main/split_files/text_encoders (already installed by anyone running Qwen Image) - Original open weights: https://huggingface.co/tencent/Hy-Nomas - Project page: https://hy-nomas.hy.tencent.com/ - Custom node: https://github.com/PedroMarinhoDev/ComfyUI-Hy-Nomas/ - Prompt skill: https://github.com/benjiyaya/hy-nomas-prompts Calliope - Local-first Agentic Idea-to-video Studio https://github.com/benjiyaya/Calliope If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech #comfyui #tencent #aivideo #worldmodel

Comments

No comments found for this post.