Rodin - Detail Generate

Before the Gen 2.5 lineup existed, Hyper3D's Rodin API had a simpler family of image-to-3D modes - Regular, Detail, Smooth, Sketch - and Rodin - Detail Generate is the high-detail member of that old club. It still ships in the ComfyUI-Rodin pack and it still works fine. It just does a different job than the newer text nodes, so don't expect a prompt box here: this one is image-driven, and that's the point.

What it is

mRodin3D_Detail takes one or more images and turns them into a detailed 3D mesh with textures. You feed it a reference image (or several views of the same object), it uploads them to the Hyper3D Rodin API, and the cloud does the reconstruction. Like every node in this pack, nothing is computed locally - no GPU needed, but you do need an API key with credits, because this is a paid service wearing a ComfyUI costume.

One thing to get straight early: the name is slightly misleading. It doesn't take an existing model and add detail to it. It's the "detail-tuned" generation mode of the Gen 1.5 line - you give it images, it produces a mesh aiming for high fidelity, and the tier="Detail" flag is hardcoded in the source.

How it works

The node runs the same process_full_generation() pipeline as its siblings: submit a job, poll the API until it's done, then download the result into a timestamped folder under ComfyUI/output/. The single output, model_path, is a string path to the downloaded file.

Two source-level details worth knowing. First, multiple images get sent as a batch - the code converts each image tensor and sends them all, which is how you get multi-view reconstruction rather than just "make a model of this one photo." Second, there's no format dropdown on this node: geometry_file_format is hardcoded to glb. You get a .glb, full stop.

The inputs that matter

• images - the whole point. An IMAGE input, so it wires straight from a Load Image node or anything upstream. One image works; several views of the same object usually give a better result.
• api_key - from the Rodin - API KEY node. No key, no run.
• Polygon_count - six options: 4K-Quad through 200K-Quad, plus 200K-Triangle. Default is 18K-Quad. The Quad options are quad-dominant meshes (better for sculpting and animation), Triangle is raw. This is your quality/size dial, and it maps directly onto the generation job.
• Material_Type - PBR, Shaded, All, or None. Note the default here is PBR, unlike the Gen 2.5 text nodes which default to Shaded. PBR gives you the full material set - worth keeping if you plan to render the model elsewhere.
• seed_ - matters if you want reproducible results across reruns.

Two optional inputs exist for control: bbox (from the Rodin - BBox Controlnet node, to force dimensions) and height_cm (physical scale). Skip them on a first pass.

Installing it

Same as every node in this pack. ComfyUI Manager, search ComfyUI-Rodin, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Ravenmelt/ComfyUI-Rodin
# restart ComfyUI

The only real Python dependency is aiohttp for the async HTTP calls (asyncio is listed but it's standard library). No model weights, no torch extras - the pack is refreshingly light compared to most of the ecosystem.

Where people get burned

Failures are cloud-shaped, same as the other Rodin nodes. Forget the key and you get MISS_API_KEY; feed no images and you get MISS_IMAGES_OR_PROMPT. RODIN_ERROR is the API rejecting your job, NO_JOBS_FOUND / MISS_SUBSCRIPTION_KEY point at a bad key or account, and a silent download failure usually means no credits, no internet, or no write permission on ComfyUI/output.

The bigger gotcha is expectations. This is a minutes-long cloud job that blocks your queue, and it costs credits per run. Iterating on a 3D model this way is slow and not cheap, so it pays to nail the input images first - clean background, consistent lighting, multiple views when you need fidelity.

My take: it's a dated node in a pack with shinier toys, but if you have a reference image and want a quad-friendly, detailed mesh without training anything, it's still the one to reach for. For pure text prompts, the Gen 2.5 text nodes are the better fit.