HYWorld2 World Expansion

A point cloud and a memory bank get you a frozen snapshot. HYWorld2 World Expansion is where the scene starts becoming a world you can move through - it runs the WorldStereo video-generation stage over your planned trajectories, generating new frames for each camera path, then merges those results back into the memory bank. This is the "generate more of the scene so it's actually explorable" step, and it's the heaviest model lift in the pipeline.

The model it uses comes from the WorldStereo family - the node takes a WORLDSTEREO_MODEL input, which you load with the pack's VNCCS_LoadWorldStereoLightModel node. (WorldStereo is the video-then-reconstruct side of HY-World; the author's own release notes flagged these models as heavy and somewhat dated.) On top of that, the node drives a Qwen3-VL model for the captioning stage, so this one node can pull two large models' worth of weights on first run.

How it works

Four stages, roughly:

1. Caption - captions each rendered trajectory with Qwen3-VL (traj_caption.json per trajectory), reusing existing captions unless forced. This used to live in Trajectories; in pack 1.5.1 it moved here.
2. Generate - runs WorldStereo video generation over the trajectory keyframes, conditioned on the memory bank so the new footage stays consistent with what's already there.
3. Merge - folds the generated results back into the memory bank.
4. Return - the updated bank plus a JSON status.

The Qwen knobs are all prefixed qwen_: qwen_model_id (default Qwen3-VL-4B-Instruct), qwen_quantization (4-bit default to fit VRAM), qwen_frame_count (how many frames of context, default 4), qwen_max_new_tokens, qwen_attention_mode. max_trajectories (0 = all) lets you cap how many paths get expanded.

The inputs that matter

• workspace, memory_bank, trajectory_set, model - all required. The model socket is WORLDSTEREO_MODEL, from the WorldStereoLight loader.
• seed - for reproducibility across runs.
• qwen_quantization - leave at 4-bit unless you have VRAM to burn.

Outputs

• memory_bank - updated with the expanded geometry; wire onward to GS Data / Train 3DGS.
• info - JSON.

Install & troubleshooting

Pack install: ComfyUI Manager (search HY-World 2.0) or clone + pip install -r requirements.txt + python install.py. WorldStereo rendering additionally needs PyTorch3D (install manually: pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable") and the MoGe depth estimator (pulled from the Microsoft GitHub repo by requirements.txt). Diffusers and transformers are in the requirements too, and they're big installs - this is the most dependency-heavy corner of the pack.

VRAM is the wall. The author tested on a 16 GB card with 64 GB of RAM and flagged that decent resolutions "require a great deal of VRAM." If generation OOMs, drop qwen_frame_count and max_trajectories first, then lower resolution. And manage expectations: HY-World2 itself is mostly WorldMirror-plus, and the full interactive world pipeline Tencent demoed isn't released yet - what this pack gives you is the gaussian-splat half, genuinely, but the "walk around and control a character" part is still not there.