Rodin Gen-2

Most of the Runware pack's nodes hand you an image or a video. This one hands you a 3D model. Rodin Gen-2 is Hyper3D's (Deemos) image-to-3D and text-to-3D model, and through this node it turns a photo or a prompt into a proper GLB mesh with PBR materials - no Blender, no photogrammetry rig, no local GPU. You get a usable asset for game prototyping, product shots, or just seeing what your favorite 2D character looks like in the round.

A couple of honest caveats before you get excited. First, the output is a file_path string, not something ComfyUI can render in the viewport - the model saves to your output folder and the node hands you the path. That's the tradeoff for putting 3D in a 2D-first tool. Second, "Gen-2" quality on a 500-poly budget is a rough model; the detail settings exist for a reason.

How it works

The node posts an image (or just a prompt) to Runware's cloud, which runs Rodin's generation pipeline and saves the result as a GLB to your ComfyUI output folder. The request is assembled from the images input, the prompt, and a pile of mesh/texture settings, and the returned file_path string points at the saved file. Because it's an API call, quality is a parameter rather than a hardware problem - crank it and pay, or keep it light.

The inputs that matter

• images (IMAGE) - the reference. Image-to-3D is where Rodin genuinely shines: feed it a product photo or a character render and it reconstructs the shape, including alpha if you keep settings.useOriginalAlpha on.
• positivePrompt - text-to-3D fallback, and useful for nudging the image result ("sneakers, detailed soles").
• settings.quality - extra-low to high. This sets the preset polygon budget, and it's the first knob you should touch.
• settings.polyCount / settings.polyCount_value - the manual override, 500 to 1,000,000. The tooltip is blunt: 150k+ recommended for real detail. That's the difference between a usable asset and a lump.
• settings.material - PBR, Shaded, or All. PBR gets you the textured, game-ready output; All returns both.
• settings.meshMode - Raw or Quad. Quad topologies rig and animate better; pick it if the model's going into a game engine or character pipeline.
• settings.hdTexture - on for clean close-ups, off to save cost.
• settings.taPose - forces a T/A-pose on humanoids, which you want if you're rigging.

There's also seed (max 65535 here), numberResults (max 4), outputFormat (GLB only - not much of a dropdown), and an advanced_json escape hatch documented for settings.addons and settings.boundingBox.

Installing

Same pack, same drill:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart ComfyUI (or ComfyUI Manager → search Runware). Needs a Runware API key via Settings, RUNWARE_API_KEY, or runware auth login.

Where people get burned

The most common stumble is expecting the node to show you the model. It won't - the file_path output is a string. Point it at a Save/Preview text node or just check your output/ folder for the GLB and open it in any 3D viewer. Next: polygon count. Default quality is medium and the default budget is far below what looks good; if your result looks melted, that's the poly budget, not the model. And riggers, don't skip taPose - retopologizing a non-T-pose humanoid by hand is nobody's idea of a fun afternoon.