☁️BizyAir Load Trellis

What it is

This is the loader half of BizyAir's Trellis pipeline - the node you drop down first to configure Microsoft's TRELLIS image-to-3D model before you actually run anything through it. If you've never touched TRELLIS before: it's the model that takes a single product photo or character render and spits out an honest-to-god 3D asset, mesh and all, no photogrammetry rig required. It's genuinely impressive when it works.

The catch with vanilla TRELLIS is that getting it running locally is its own small nightmare - CUDA versions, Visual Studio build tools for the spconv compilation, a specific PyTorch pin, the works. People who've fought through it describe multiple clean OS reinstalls just to get the dependencies to agree with each other. BizyAir's whole pitch is skipping that: this loader just points at TRELLIS running on their infrastructure, so you pick your settings and let their GPUs deal with the compilation headaches.

The inputs that matter

Most of these are performance/quality knobs rather than creative ones, so you can leave the defaults alone for your first few runs:

• model_name - currently just TRELLIS-image-large, the standard checkpoint.
• dinov2_model - which DINOv2 vision backbone reads your input image. The default (dinov2_vitl14_reg) is fine; the larger dinov2_vitg14_reg option trades speed for slightly better results per the node's own tooltip.
• use_fp16 - half-precision, on by default. Leave it on.
• attn_backend - which attention kernel does the heavy lifting: sage (default, described as experimental but faster), xformers, flash_attn, sdpa, or naive. This is exactly the kind of setting that's a pain to get working on your own machine and a non-issue here.
• smooth_k / spconv_algo - sparse-convolution tuning. implicit_gemm (the default spconv algo) is described as the most accurate but slower option, with native being the fast-and-loose alternative. Not something you'll normally touch.

The single output is a model handle of type TRELLIS_MODEL - wire that into BizyAir Trellis Predict (BizyAir_IF_TrellisImageTo3D), which is the node that actually does the image-to-3D conversion.

How to install it

Easiest path is ComfyUI Manager: open it, search "BizyAir," hit install, restart. If you'd rather do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

Then restart ComfyUI. There's also a comfy node install bizyair route if you're using the Comfy CLI.

Before anything runs, you need a BizyAir account and API key set in the node's settings - this whole pack is a thin client for SiliconFlow's cloud, and without a key every node in it just fails at the first execution.

Common issues

"Missing node" installs BizyAir without you asking for it. This has bitten a bunch of people: ComfyUI Manager's older node-registry inference occasionally misattributed missing nodes from other packs (Nunchaku's Flux workflows, notably) to BizyAir, so people would go to fix an unrelated missing-node error and suddenly have an unfamiliar folder with Chinese branding sitting in their custom_nodes directory. It reads as sketchy the first time you see it. It isn't malware - SiliconFlow is a real commercial inference company - but ComfyUI's own maintainers had to patch the registry mapping in mid-2025 because the false-attribution was common enough to spook people. If you installed this loader on purpose, ignore all that; if you find a BizyAir folder you didn't ask for, it's safe to just uninstall it from the Manager.

Holes in the generated mesh. This is a known TRELLIS quirk more than a BizyAir bug - people running TRELLIS locally hit the same gaps in the output geometry. It tends to show up on thin or occluded parts of the subject; a cleaner, well-lit reference photo with the object fully in frame is the most reliable fix.

No TRELLIS_MODEL output / downstream node errors. Double-check your API key is set - a failed auth call here silently breaks the whole chain, since the Predict node has nothing valid to work with.