Hy-Nomas x LTX-2.5 Audio Workflows
Tencent's Hy-Nomas is an 8B video world model: you write a world, it renders it segment by segment, and you can keep extending it or drive it with WASD. It has no audio.
These two ComfyUI workflows give it a soundtrack with the stock LTX-2.5 distilled model. No training, no new models. LTX-2.5 is a joint audio-video model, so I freeze the Hy-Nomas frames inside LTX's latent and let it generate only the audio. Because the audio attends to the frozen video, the sound follows the picture: hoofbeats while the horse gallops, a deep burner roar the moment the flame fires.
Examples (straight out of the workflows)
Hot air balloon over Giza: Text to World with Audio, 1280x704, 10 s, seed 7 (the workflow ships with 42; for this one set Sample segment 1 to 7, Sample segment 2 to 8 and the LTX Seed to 7)
Calm breeze over the desert, then the deep burner roar lands right as the view tilts up to the flame.
Dragon raid: Add Audio to Any Video, on a Hy-Nomas clip, 1280x704, 10 s, seed 42
Regular hoofbeats through the gallop, then a broadband blast exactly when the wall of fire erupts. With the video switched off (same prompt, LTX's audio-only mode) you just get one flat roar from frame 0, so the frames really are steering it.
World prompts and sound prompts for both are in examples/.
The workflows
1. Add Audio to Any Video: workflows/hynomas_x_ltx25_add_audio_to_video.json
Core ComfyUI nodes only. Load a Hy-Nomas world, a saved Hy-Nomas Drive, or any clip, write a sound prompt, run. The video is saved untouched with the new audio track.
2. Text to World with Audio: workflows/hynomas_x_ltx25_text_to_world_with_audio.json
Writes a 2-segment Hy-Nomas world (stock T8 settings: 8 steps, euler, simple, CFG 1) and scores it with LTX-2.5 in one run. Ships with the hot air balloon world as the default.
How it works
- Pad the clip to LTX's frame grid (8n+1) by holding the last frame, all in-graph, so any length works.
- Encode the frames with LTX's own video VAE at half size (the same size the stock LTX template's first stage uses).
- Freeze that video latent (
LTXV Freeze Latent, noise mask 0) so LTX cannot change a single frame. - Join it with an empty LTX audio latent and sample with the stock distilled sigmas (euler ancestral, CFG 1).
- Decode with the LTX audio VAE and mux it onto the original full-resolution frames.
Speed on my RTX 5090 (32 GB, ComfyUI 0.44.0)
| Time | Peak VRAM | |
|---|---|---|
| Add Audio to a 10 s 1280x704 clip | ~20 s warm, 41 s with first model load | ~31 GiB |
| Text to World with Audio, 10 s at 1280x704 | 103 s (132 s with first model loads) | ~31 GiB |
No offload in any run.
RAM: Text to World loads Hy-Nomas and LTX-2.5 in one run and peaked at about 62 GB of system RAM on my 64 GB box. It works, but it's at the limit. With 64 GB, close other apps, or render the world first and score it with workflow 1 in a second run.
What you need
- ComfyUI 0.43 or newer (tested on 0.44.0).
- LTX-2.5 files from Lightricks/LTX-2.5 (gated, accept the license first):
diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensorstext_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensorsvae/ltx-2.5-video-vae-bf16.safetensorsvae/ltx-2.5-audio-vae-bf16.safetensors
- Workflow 2 only: the Comfyui-HY-Nomas-T8 node pack (ComfyUI-Manager: "Hy-Nomas T8") and its three files from t8star/HY-Nomas-Comfy:
hy_nomas_distil_int8.safetensors(diffusion_models),hy_nomas_pixelumm_fp8.safetensors(vae),qwen_2.5_vl_7b_fp8_scaled.safetensors(text_encoders).
Tips
- Write the sound prompt like an LTX prompt: one flowing paragraph, describe the scene briefly, then every sound and when it happens. End with "with no music" if you don't want a score.
- The raw audio comes out hot (around -9 LUFS). Normalize it in your editor if you need broadcast levels.
- Got a Hy-Nomas Drive you explored with WASD? Its Save mp4 lands in
output/hy_nomas/. Load that into workflow 1 and it gets a soundtrack too.
Licenses
These workflow files: MIT. The models keep their own licenses: LTX-2.5 under the LTX-2.x Community License, the Hy-Nomas DiT under Apache-2.0, and the PixelUMM decoder under NVIDIA's One-Way Noncommercial license (research and evaluation only). This repo ships no model weights.
Credits
Hy-Nomas by Tencent Hunyuan. ComfyUI nodes by T8star, built on PedroMarinhoDev's ComfyUI-Hy-Nomas. LTX-2.5 by Lightricks.