Hy-Nomas x LTX-2.5 Audio Workflows

Hy-Nomas x LTX-2.5 Audio Workflows

Two ComfyUI workflows that add a soundtrack to silent Hy-Nomas video worlds or any clip, using the stock LTX-2.5 distilled model while the picture stays untouched.

Made by BennyDaBall on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: These two ComfyUI workflows add a soundtrack to video from Hy-Nomas, Tencent's 8B video world model, which renders silent clips. They run the stock LTX-2.5 distilled model, so nothing is trained and no extra models are needed. The video frames stay frozen while LTX generates only the audio, which means the sound follows what happens on screen.

Who it suits: It suits people who already make video worlds with Hy-Nomas in ComfyUI and want sound without leaving the graph. It also takes any other clip, so editors with silent footage can use it too.

Industry | Who uses it | How they use it | Example
Film and video | Video artists | Add sound to a silent Hy-Nomas world clip | A galloping horse clip gets hoofbeats that match the motion
Games and worlds | People who explore worlds with WASD | Score a saved Hy-Nomas Drive | The saved mp4 loads into workflow 1 and gains a soundtrack
Content creation | Makers of short videos | Put audio on a clip they already have | The clip comes back with the same frames and a new audio track

You can skip it if you do not run ComfyUI. The author measured a peak near 31 GiB of VRAM with no offload, so check your graphics card first.

Where to use it:

• ComfyUI 0.43 or newer, where the author tested 0.44.0
• Windows, macOS or Linux with Python, to download the files
• Hy-Nomas world clips and saved Hy-Nomas Drive videos
• Any other video clip you load into the first workflow

Use case | Who | What they get back
Add sound to an existing clip | Video makers | The same video with a new audio track
Write a world and score it in one run | Hy-Nomas users | A two-segment world with LTX-generated audio
Soundtrack a WASD exploration | World explorers | A saved mp4 that now has sound

How to set it up: You need a working ComfyUI and some model files before you start.

• ComfyUI 0.43 or newer
• Four LTX-2.5 files from the Lightricks repository on Hugging Face, which is gated, so accept its license first
• Workflow 2 only: the Hy-Nomas node pack from ComfyUI-Manager, plus three Hy-Nomas files from the t8star repository

1. Install the Hugging Face tools with Python.
2. Download this repository with the second command below.
3. Copy the four LTX-2.5 files into the diffusion_models, text_encoders and vae folders of ComfyUI.
4. For workflow 2, install the node pack and add its three files to the matching folders.
5. Open a workflow JSON file from the workflows folder in ComfyUI.

pip install -U huggingface_hub
hf download BennyDaBall/Hy-Nomas-x-LTX-2.5-Audio-Workflows

The four LTX-2.5 files sit at these paths:

diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors
text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors
vae/ltx-2.5-video-vae-bf16.safetensors
vae/ltx-2.5-audio-vae-bf16.safetensors

Workflow 2 also needs these three Hy-Nomas files:

hy_nomas_distil_int8.safetensors
hy_nomas_pixelumm_fp8.safetensors
qwen_2.5_vl_7b_fp8_scaled.safetensors

How to use it: Load a workflow and type a sound prompt, then run it. Keep the prompt to a single paragraph in the style of an LTX prompt. Open with a short description of the scene, then name each sound and the moment it happens. Finish with the words with no music if you do not want a score.

Tool | What it does
Add Audio to Any Video | Takes a Hy-Nomas world, a saved Drive or any clip plus your sound prompt, and returns the same frames with a new audio track
Text to World with Audio | Writes a two-segment Hy-Nomas world with stock T8 settings, then adds LTX-2.5 audio in a single run

The audio comes out hot at around -9 LUFS, so bring it down in an editor if you need broadcast levels. A Drive you explored with WASD saves its mp4 into output/hy_nomas, and workflow 1 accepts that file.

Example: You load a 10 s Hy-Nomas clip of a dragon raid at 1280x704 into the first workflow and keep seed 42. Hoofbeats keep a steady beat through the gallop, then a broadband blast hits at the moment the wall of fire erupts. Switch the video off and the same prompt gives one flat roar from the first frame, so the frames do steer the sound.

For newcomers and experts: Think of it as a sound designer for silent AI video. Start with workflow 1: load one clip, type a short description of the sounds you want, and run it. The clip comes back with an audio track and an unchanged picture.

Under the hood, the clip is padded to LTX's frame grid of 8n+1 by repeating the last frame, then encoded with LTX's video VAE at half size. The video latent is frozen with a noise mask of 0, joined with an empty audio latent and sampled with the stock distilled sigmas, euler ancestral at CFG 1. The audio is generated while the video is locked in place, so its timing follows the picture. The trade-off is memory, since workflow 2 loads two large models in one run.

Pad the clip to the 8n+1 frame grid
Encode the frames with the LTX video VAE
Freeze the video latent
Sample the audio with distilled sigmas
Decode the audio and mux it onto the frames

Good to know:

• Speed: on an RTX 5090 with 32 GB, a 10 s clip took about 20 s warm in workflow 1 and 103 s in workflow 2
• Memory: peak VRAM was near 31 GiB with no offload, and workflow 2 peaked near 62 GB of system RAM on a 64 GB machine
• Tight on RAM: render the world in one run, and let workflow 1 add the audio in a second
• License: MIT
• Cost to run: none named, but the LTX-2.5 files are gated
• Data it reaches: the clips you load, and it writes the new video with audio to ComfyUI's output folder
• Not for: changing the picture, since only the audio is generated

Title: Run time on an RTX 5090
Add Audio, warm: 20 s
Add Audio, first load: 41 s
Text to World: 103 s
Text to World, first load: 132 s

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 4
• License: mit
• Tags: comfyui, workflow, hy-nomas, ltx-2.5, video-to-audio, world-model, audio-generation

Links

• Model: https://huggingface.co/BennyDaBall/Hy-Nomas-x-LTX-2.5-Audio-Workflows
• Files: https://huggingface.co/BennyDaBall/Hy-Nomas-x-LTX-2.5-Audio-Workflows/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BennyDaBall/Hy-Nomas-x-LTX-2.5-Audio-Workflows

Then follow the usage in its model card. Nothing in this kit runs by itself.