HunyuanWorld Image to Panorama

You have one photo of a place - a friend's apartment, a location scout still, a render you like - and you want the rest of the universe around it. That's this node. HYWorldImage2Pano takes a single perspective image, figures out what's behind and to the sides, and hallucinates the missing 360° into a seam-free equirectangular panorama. It's the image-in variant of the pack's text-in panorama node, and it's the more impressive trick of the two: given a decent source, the results feel like someone handed you a 360° camera that was pointed at your scene.

How it works

The heavy lifting is a FLUX.1-Fill-dev inpainting pipeline (black-forest-labs/FLUX.1-Fill-dev, also gated, also ~23 GB) with HunyuanWorld's PanoDiT-Image LoRA fused in. The clever part is the geometry, which the node does for you: your source image is resized to match the fov you declare, projected into equirectangular space (that's the Perspective → GetEquirec step), and the empty region is turned into an inpaint mask. The model then fills everything the mask covers - and it has to match the source's perspective, lighting, and style, which is exactly why an inpainting-tuned FLUX is the base. It also quietly appends its own quality prompt ("high-quality, high-resolution, sharp, 8k") and a negative that bans people and mess, so you don't have to.

Inputs and outputs that matter

• image - required. Any IMAGE tensor you can produce: Load Image, a render, a screenshot.
• fov - default 80, range 30–180. This is the one to get right. It tells the node how wide your source photo is, and it drives the whole projection. A typical phone photo is roughly 65–75°; a game screenshot with a narrow lens might be 50. Say 120 when your lens is actually 70 and the model will warp the source before it even starts.
• prompt - optional text guidance. Handy for steering what the model fills in ("dense forest continues," "city skyline"). Empty is fine.
• height / width - panorama size, default 960×1920. Height tops out at 2048, width at 4096.
• guidance_scale - default 30 again, the HunyuanWorld house style. Don't dial it to 7 because FLUX told you so.
• blend_extend, fp8_quantization, use_cache - the seam blending and VRAM knobs, same as the text node.

Outputs: panorama (IMAGE) and panorama_path (STRING). Keep the string - HYWorldSceneGen wants a file path to feed the 3D stage.

Installing it

Same pack install as the rest: ComfyUI Manager → HunyuanWorld 3D World Generation, or

cd ComfyUI/custom_nodes
git clone https://github.com/krmahil/comfyui-hunyuan-world.git
pip install -r comfyui-hunyuan-world/requirements.txt

Then the pack prerequisites: HunyuanWorld-1.0 installed so hy3dworld imports, Real-ESRGAN, the ZIM segmentation models, and a HuggingFace login with the FLUX.1-Fill-dev license accepted. First run downloads the Fill model plus the ~478 MB PanoDiT-Image LoRA - budget 10–30 minutes.

Common issues

• The source image looks warped - fov is wrong for your source. Measure or guess tighter before blaming the model.
• Seam or lighting mismatch - nudge blend_extend up, and give the prompt a hint about the scene continuing.
• Gated model errors - accept FLUX.1-Fill-dev on HuggingFace and log in; it's a different gated repo from FLUX.1-dev, so accepting one doesn't grant the other.
• OOM - same story as the text node: FP8 (fp8_quantization) gets you to ~24 GB, adding use_cache (DeepCache) to ~20 GB. Without them you need ~40 GB.

Both FLUX pipelines (dev and Fill-dev) stay cached in memory after their first use - if you run this node and the text node in the same session, that's ~46 GB of models parked in VRAM. Throw a HYWorldUnloadModels at the end of the graph when you're done.