Idol for ComfyUI

Idol (Speridlabs, Apache-2.0) is a single image to 3D gaussian splat model. These are its released weights in the file layout ComfyUI's stock loaders take, for ComfyUI-Idol. Idol is the TripoSplat architecture, which ComfyUI implements natively, so no model code is needed: the files load with Load Diffusion Model and Load VAE and run on the stock TripoSplat nodes.

File	ComfyUI folder	Size	Contents
diffusion_models/idol_fp16.safetensors	models/diffusion_models	0.74 GB	the denoiser (denoiser.safetensors), fp16
vae/idol_decoder_fp16.safetensors	models/vae	0.58 GB	octree + gaussian decoders (decoder.safetensors + decoder_gs.safetensors) in one file, fp16

Also needed, from where they are already published: dino_v3_vit_h.safetensors and flux2-vae.safetensors (VAST-AI/TripoSplat) and birefnet.safetensors (Comfy-Org/BiRefNet). Idol's own birefnet.safetensors is that file without its training heads.

What changed

Made with scripts/convert.py from speridlabs/idol; parameter names are unchanged.

• Denoiser: cast to fp16 (the precision the reference runs the transformer trunk at). The pos_pe buffer (the latent tokens' fixed Sobol positions) is dropped because ComfyUI recomputes the identical sequence.
• Decoders: the two files are merged under the octree. and gs. prefixes ComfyUI detects the TripoSplat gaussian decoder by, cast to fp16. The stored perturbation buffers are kept.

From the fp32 weights ComfyUI's path matches the reference implementation to 3e-6 relative error for the denoiser and ~1e-4 to 3e-3 for the decoders (ComfyUI's RMSNorm eps); these fp16 files add ~1e-3 (tests/parity.py in the repo).

Usage

See the ComfyUI-Idol README: KSampler euler / simple, 50 steps, cfg 4.0 (Idol's cfg_strength 3 and time shift 3), TripoSplat Decode with 262144 gaussians.

License

Apache-2.0, as the original weights. Idol by Speridlabs.