HY-World Qwen Pano Encoder

Before Qwen can expand your image into a 360° panorama, someone has to lay down the blank canvas it's going to expand onto, and build the prompt/conditioning that tells it what to do. That's HYWorld2 Qwen Pano Encoder. It takes your input image, crops and resizes it into a reference, creates a wide empty canvas the model will fill in, and emits the positive/negative prompt pair - everything the Qwen Pano Generate node needs to start sampling.

Think of it as the "set the table" node in a two-course meal: Encoder prepares, Generate cooks. You don't strictly need it if you'd rather hand-build the canvas yourself, but it's the intended path, and it's where you set the output resolution for your panorama.

How it works

From the source, the mechanism is refreshingly concrete:

1. It takes your image, crops off the crop_border fraction from the edges (default 0 - the tooltip-relevant knob is there for pano sources that have lens/edge artifacts).
2. It builds a blank canvas tensor of exactly width × height (defaults 1952×960, the equirect aspect the Qwen pano pipeline expects), filled with canvas_value (0 = black). batch_size tiles the reference across multiple canvases if you're generating a batch.
3. It constructs the prompts - same official template logic as Qwen Pano Prompt, with use_official_prompt on by default and your scene_prompt describing the continuation.
4. It returns the reference_image (cropped), the wide_canvas_image (blank), and the two prompt strings.

The inputs that matter

• image - required. Your seed photo.
• width / height - the output panorama size. 1952×960 is the pipeline default and a good start; both step by 8.
• scene_prompt - what the expansion should depict in the extended areas.
• crop_border - nudge up (max 0.49) if your source has edge junk that would get hallucinated into the panorama.
• canvas_value - black is right for this model family; you can lighten it but expect the model to fight you.

Outputs

• reference_image and wide_canvas_image - the two images to feed Qwen Pano Generate (it takes the reference; the canvas informs the empty latent).
• positive_prompt / negative_prompt - the conditioned strings for the generator.

Install & troubleshooting

Pack install: ComfyUI Manager (search HY-World 2.0) or clone + pip install -r requirements.txt + python install.py. No model downloads here - this node is pure tensor wrangling, so it's fast and light.

The subtle stuff is all in the resolution: keep width/height multiples of 8 (the node enforces the step), and remember the pipeline was tuned at 1952×960 - going wildly wider works but costs VRAM at generation time. If your reference keeps getting cropped unexpectedly, you left crop_border above 0 from a previous experiment. And because this feeds a generative model, garbage in = pano garbage out: a sharp, cleanly-cropped input is the single biggest quality lever you control at this node.