← sources

Hy-Nomas (Tencent's video world model) running natively in ComfyUI with just 12GB VRAM

source: reddit (reddit:StableDiffusion)published 2026-10-23 22:32ingested 2026-10-24 02:38

retained text

Tencent released [Hy-Nomas](https://huggingface.co/tencent/Hy-Nomas), a video world model: it generates a world a few seconds at a time, and every new segment remembers the earlier ones (it caches their keys/values), so you can keep going and change what happens as you go. I ported it to native ComfyUI nodes: the DiT is a normal ComfyUI MODEL sampled by the stock KSampler, with ComfyUI's own memory management.

**What it does**

* **Start from your own image.** The model has no image input, so the node encodes your picture as a short still clip and puts it in the model's memory as the first segment; the world continues from it. Not something Tencent trained, but the look carries over well (the Paris and beach clips in the video start from pictures).
* **Chain segments**, each with its own text: actions, scene changes, camera moves.
* **Camera control**: key timelines ("D+Left 3" circles the subject), or fly a path in a 3-D recorder inside the node.
* **Drive mode**: hold WASD and the arrows in the node; it generates 1 s turns and streams the frames back. Not real time: about 4.3 s per 1 s turn at 832x480 on a 4090.

**Speed** (RTX 4090): 1.02 s per 720p sampling step (a plain bf16 port takes 2.29 s): Tencent's INT8 weights run on the int8 tensor cores through a Triton block-scaled GEMM, plus fused kernels, an FP8 decoder and int8 memory in RAM. 10 s of 832x480 takes 34 s end to end; a 20 s 720p world (five chained segments) about 3 minutes. On 12 GB and 16 GB (simulated by capping a 4090) it runs at nearly the same speed: 93 s vs 89 s for 10 s of 720p.

**A tip if you try it**: write the world prompt in the format Tencent's prompt rewriter produces (`<|Subject1|`, `<Action1...</Action1`, follow-ups starting "Compared to the previous segment, "). Written loosely, follow-up segments are mostly ignored. The README explains it, and the example workflows use it.

**Work in progress**: it works end to end and is checked block by block against Tencent's code, but expect rough edges. Planned: Tencent's other checkpoints and LoRAs, a faster decoder, a closer-to-real-time drive mode, a prompt helper.

* Nodes: [https://github.com/PedroMarinhoDev/ComfyUI-Hy-Nomas](https://github.com/PedroMarinhoDev/ComfyUI-Hy-Nomas)
* Weights: [https://huggingface.co/PedroMarinhoDev/Hy-Nomas-ComfyUI](https://huggingface.co/PedroMarinhoDev/Hy-Nomas-ComfyUI)

document fetched 2026-10-24 10:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xds59e/hynomas_tencents_video_world_model_running