Note: This thing kinda sucks compared to trellis. The workflow sat in my drafts for a while but I figured I better get it out there because somebody might find it useful. I probably wont provide much support for this one unless I'm doing something dumb in the workflow or if there is a way to make this thing better.

Type the name of any object in a picture and it turns into 3D. You get a .ply file you can turn around, and a spin video if you want one.

Most picture-to-3D workflows want a clean product shot. This one starts from any photo: type the thing you want, like the camera, and only that goes to Idol.

How to use it

1. Load your picture.

2. In Type what you want, type the thing to take out, like the camera.

3. Press Run. What it found shows what was picked.

4. Drag inside Save Splat to turn your object around.

Switches

• Fill in hidden parts draws what a hand covers. Off at the start. Leave it off for people and statues. Big download, about 17 GB.

• Sharpen with SeedVR2 adds detail before the build. On at the start.

• 1 million splats for a denser, bigger file.

• Looser framing if the edges get cut off.

• Save a spin video makes a 3 second loop, seen straight on at eye level. Click background to pick its color.

• Make a mesh (GLB) for Blender or a game engine.

What you need

• AusBoss nodes 2.21.2 or newer (update first), ComfyUI 0.43.2 or newer. No other node packs.

• idol_fp16.safetensors → models/diffusion_models · 741 MB

• idol_decoder_fp16.safetensors → models/vae · 576 MB

• dino_v3_vit_h_bf16.safetensors → models/clip_vision · 1.68 GB

• flux2-vae.safetensors → models/vae · 336 MB

• sam3.1_multiplex_fp16.safetensors → models/checkpoints · 1.75 GB

• For Sharpen with SeedVR2 (on at the start): seedvr2_3b_int8_convrot.safetensors → models/diffusion_models · 3.46 GB

• For Sharpen with SeedVR2: seedvr2_ema_vae_fp16.safetensors → models/vae · 501 MB

• Only for Fill in hidden parts (off at the start, about 17 GB in all): qwen_image_2.1_int8_convrot.safetensors → models/diffusion_models · 7.26 GB

• Only for Fill in hidden parts: qwen3vl_8b_int8_convrot.safetensors → models/text_encoders · 9.35 GB

• Only for Fill in hidden parts: qwen_image_2.1_vae_bf16.safetensors → models/vae · 676 MB

• Only for Fill in hidden parts, rename model.safetensors to this name: Qwen-Image-2.1-LayerExtract.safetensors → models/loras · 168 MB

Tips

• Works best on a clear, front-on photo of one object, in even light.

• Idol guesses the depth from one picture, so some shapes come out squashed. Try another photo.

• Say it more clearly if it picks the wrong thing: the red umbrella, not umbrella.

• The back and sides are made up to match the front. Change the Seed and run again for a different guess.

• It is built for one object. People and rooms don't come out well.

• In the 3D viewer, scroll to zoom in. It opens small and far away.

Credits

Idol is by speridlabs (Apache-2.0), with ComfyUI files from Comfy-Org. DINOv3 and SAM 3 are by Meta, the FLUX.2 VAE by Black Forest Labs, SeedVR2 by ByteDance Seed, Qwen Image 2.1 by Qwen, and Layer Extract by DiffSynth-Studio. Each has its own license. Built with DINOv3.

v1.0 (2026-11-03): first release.

Questions and bugs: GitHub issues. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.