Home Creators Posts Import Register Favorites Logout

Videos

  • 777695128.mp4

Downloads

Content

Hey everyone.

If you watched the Hy-Nomas video

https://www.patreon.com/aifuturetech/posts/hy-nomas-video-174376205/

https://youtu.be/JMDPignEbP0 .

You saw the full setup and the demos. This post has the two advanced workflows as runnable files: the 40-second four-segment world and the camera-control build. These are the two I picked from everything in the video because they are the most practical ComfyUI use cases: one gives you a long one-shot video from a JSON script, the other lets you fly the camera yourself and have the world follow. Everything below is parsed straight from the workflow files, so what you read here is what you download.

----------------------------------------

WHAT TO DOWNLOAD (MODEL FILES)

----------------------------------------

Custom node first: ComfyUI-HyNomas (install from ComfyUI Manager, or github.com/PedroMarinhoDev/ComfyUI-Hy-Nomas).

Then three model files. Rename freely, the loaders just need to find them:

- models/diffusion_models: hy_nomas_distil_int8.safetensors (the INT8 distilled DiT)

Original file: DiT/distilled_int8/diffusion_pytorch_model_ema_0_995_int8.safetensors

from https://huggingface.co/tencent/Hy-Nomas (also mirrored in PedroMarinhoDev/Hy-Nomas-ComfyUI)

- models/vae: hy_nomas_pixelumm_fp8.safetensors (the FP8 PixelUMM decoder)

Original file: VAE/PixelUMM_FT_FP8/diffusion_pytorch_model_fp8.safetensors from tencent/Hy-Nomas

- models/text_encoders: qwen_2.5_vl_7b_fp8_scaled.safetensors

from https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI

If you already run Qwen Image, you have this file. In the CLIP Loader, type must be qwen_image.

License note: the PixelUMM decoder weights are under NVIDIA's One-Way Noncommercial license. Research and evaluation use only, no commercial outputs.

----------------------------------------

QUICK START

----------------------------------------

Workflow 1: "Hy-Nomas world 4 segments - 40 seconds.json" (the long one-shot video)

1. Drop the three model files into the folders above and restart ComfyUI.

2. Load the workflow. Check the three loaders (DiT, decoder, CLIP) point at your files.

3. Open the World Prompt node and paste or edit your JSON world (format below).

4. Keep width and height identical in all four Segment nodes (shipped at 1280x704).

5. Queue. Each segment samples, extends the world, and the final Decode node turns the whole thing into one 40-second clip.

That's the whole loop. The VAE decode at the end is the slow part, that is normal.

Workflow 2: "Hy-Nomas camera control.json" (you fly the camera)

1. Load the workflow and check the loaders.

2. Write your World Prompt, but do NOT describe camera moves in the text. The camera nodes own the motion.

3. Double-click the Camera Recording node. Press Record, then hold W/A/S/D to move, arrow keys (or drag the view) to turn, Shift to run. Press Stop when done, Play to check it.

4. Segment 1 uses Camera Keys instead: the same controls written as a timeline, entries like "Left 2.5, W 2.5".

5. Queue. The recorded path is stretched over its segment; only the shape of the path matters.

----------------------------------------

WHAT TO INPUT (PROMPT FORMAT)

----------------------------------------

Hy-Nomas does not take loose prose. It takes a World Prompt JSON, and loosely written follow-up segments are mostly ignored. The skeleton every prompt follows:

{

"perspective_view": "first-person",

"Style": "real-world footage",

"subject_list": [

{

"subject_id": "<|Subject1|>",

"display_name": "first-person IG influencer",

"subject_attribute": "A young woman holding a smartphone at arm's length for a selfie video..."

}

],

"segments": [

{ "start_time": ..., "end_time": ..., "actions": "..." }

]

}

The rules that actually matter:

- The subject is written as <|Subject1|>, and actions wrap it like <Action1><|Subject1|> ...your action... </Action1>.

- Later segments should start with "Compared to the previous segment, ...". This is how the model was trained to continue.

- Segment lengths come from start_time and end_time, not from any frame count. At 24 fps, four 10-second segments give you the 40-second video.

- The shipped 40-second workflow uses a first-person IG influencer walking from a shopping mall to a park. The camera-control workflow uses a scout spaceship in CG style as the demo world.

You do not have to write these by hand.

My hy-nomas-prompts skill (github.com/benjiyaya/hy-nomas-prompts) generates the full JSON from a plain idea plus a total duration, which is exactly how the 40-second mall-to-park script in the video was made. Load it on top of your agent, describe the idea, ask for N seconds, paste the output into the World Prompt node.

----------------------------------------

KEY SETTINGS (AND WHY)

----------------------------------------

- KSampler: euler, simple, 8 steps, CFG 1, denoise 1. This is the distilled model's contract, leave it as shipped.

- Width and height: multiples of 32, shipped at 1280x704 (Tencent's 720p). 832x480 if your card struggles. Must be identical in every Segment node, otherwise the world breaks between segments.

- segment_index: 0, 1, 2, 3 in order. Each Segment / KSampler / Extend World trio gets its own index; the trio is one segment.

- Seeds: different seed per segment (the workflows ship with randomize). Do not reuse one seed across segments.

- History frames: 18 as shipped. In the video I tested 24 to see if motion and coherence improve; feel free to try both, 18 is the safe default.

- Override text (per-segment prompt override): instead of running the default script for a segment, you can feed that segment a different text instruction. The number of seconds stays controlled by the JSON times, never by the node.

- Decode node: leave as shipped. One decode at the very end turns the whole world into the single output clip.

- 24 fps fixed. Duration is always start_time to end_time in the JSON.

----------------------------------------

TIPS

----------------------------------------

1. Write follow-ups as "Compared to the previous segment, ..." or they get ignored. This is the single biggest reason a multi-segment world looks like it restarted.

2. Generate the World Prompt JSON with the hy-nomas-prompts skill instead of hand-writing it. Fewer syntax mistakes, and the timings come out consistent.

3. For the camera workflow, keep camera language out of the World Prompt entirely. Camera nodes and text camera moves fight each other.

4. Record a rough camera path first, check it with Play, then queue. Only the shape of the path matters, so you do not need pixel-perfect timing while recording.

5. Leave a segment's camera unconnected if you want text-only control for that segment. Mixing is fine.

6. Drive Interactive is for fun, not production. It re-runs the whole graph every second to refresh frames, so memory fills up and colors degrade on a 24 GB card. The pre-recorded path in this workflow is the usable version of the same idea.

7. Start at 832x480 while dialing in a new world, then move to 1280x704 for the final render. Sampling is fast; the VAE decode at the end is what takes the time.

----------------------------------------

COMMON ISSUES

----------------------------------------

Second segment ignores my prompt → it is not written as a continuation. Start it with "Compared to the previous segment," and keep the action tags format.

World looks restarted or inconsistent after segment 1 → width and height differ between Segment nodes, or a seed got reused. Same resolution everywhere, fresh seed per segment.

Camera recording does nothing → the path is stretched over the segment from its times, so a 2-second recording on a 10-second segment still works; if nothing moves at all, the Recording node has no frames (press Record before moving, Stop after) or the camera input is not wired into that Segment.

CLIP loader error on startup → the text encoder file is qwen_2.5_vl_7b_fp8_scaled.safetensors and the loader type must be qwen_image, not the default.

Decode step seems stuck → it is decoding the entire world into one clip; the more segments, the longer it takes. Wait it out before re-queueing.

Colors wash out and it gets laggy → you are in Drive Interactive mode. That mode is real-time only in Tencent's web demo; locally, use the pre-recorded camera path instead.

----------------------------------------

The two workflow files and screenshots are attached. Download the three model files first if you have not run Hy-Nomas yet.

All some JSON text example prompts to run this world model.

Questions? Drop a comment with your GPU, VRAM, and the exact error, and I will help you sort it out.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji

Files

Comments

No comments found for this post.