HYWorld2 Trajectories

Once your HYWorld2 Workspace exists, HYWorld2 Trajectories decides what the camera should actually look at. This is the planning and rendering stage of the HY-World2 pipeline: it works out a set of camera paths (trajectories) through the scene, renders them into short videos, and hands the bundle to Memory Bank downstream. If you imagine the whole pipeline as "decide where to look, remember what you saw, expand the world, train a splat", this node is step one of that chain.

It's also where a vision-language model enters the picture. The node uses a Qwen3-VL model by default to caption and route-plan, which is a load-bearing part of the mechanism - the Qwen-Image/Qwen-VL family is Alibaba's, and here it's doing the "what's interesting in this scene" thinking. First run will pull the model from Hugging Face.

How it works

From the source, this node drives the worldgen trajectory generation and rendering: it builds the base geometry for the scene, plans routes (the trajectory planner relies on the recast navmesh wheel that install.py builds from the vendored HY-World source), then renders the scene's global point cloud into trajectory videos and masks under render_results/. As of pack version 1.5.1, caption generation is deferred to the downstream world-expansion stage - Trajectories just prepares and sorts render outputs, which keeps caching sane. The output trajectory_set carries the list of rendered paths and counts that later nodes iterate over.

The inputs that matter

• workspace - required, from HYWorld2 Workspace.
• scene_type - auto / indoor / outdoor. Auto is usually fine; forcing it helps the planner when the VLM misreads the scene.
• additional_nav_traj and extreme_detail_traj - turn these on when the base capture missed areas or you want dense detail coverage. detail_object_limit caps how many object-focus routes get added.
• qwen_model_id and qwen_quantization - the VLM behind the planning. Default is Qwen3-VL-4B-Instruct at 4-bit, which is the VRAM-friendly choice and honestly the one to keep.
• apply_anchor_scan / anchor_scan_topk - newer anchor-scan options; leave off unless you know you need them.

Outputs

• trajectory_set (HYWORLD2_TRAJECTORY_SET) - the rendered trajectory list, wired into HYWorld2 Memory Bank.
• info - JSON with trajectory counts and render paths.

Install & gotchas

The pack itself installs via ComfyUI Manager (search HY-World 2.0) or git clone + pip install -r requirements.txt + python install.py. The install script builds the gsplat fork and the recast navmesh wheel; if trajectory planning errors with something about navmesh or recast, install.py didn't finish - re-run scripts/pipinstall.bat (Windows) or python scripts/build_gsplat.py.

This node can be genuinely slow the first time - it renders trajectory videos, not just numbers. And if a downstream Memory Bank throws "requires completed HYWorld2 Trajectories base geometry", it means Trajectories hasn't run (or its cache was invalidated) - let it execute before the bank. Also worth knowing from the author's own release thread: this whole pack wants a chunky GPU. It was tested on a 16 GB 5070 Ti with 64 GB of RAM, and VRAM is the wall you'll hit first at high resolutions.