Idol Turbo by speridlabs

Idol Turbo by speridlabs

Turns one picture of an object into up to a million 3D Gaussians in 5 guidance-free steps, as a distilled and faster version of Idol. For 3D artists and ComfyUI users.

Made by speridlabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Idol Turbo is a distilled version of Idol, an image-to-3D model. You give it one picture of an object and it returns up to a million 3D Gaussians. It needs 5 sampling steps without guidance, where the original uses 50 guided steps, so a result takes about 0.75 s on an H200.

Who it suits: It suits 3D artists, ComfyUI users and researchers who want quick image-to-3D results and accept a small quality cost.

Industry | Who uses it | How they use it | Example
3D and games | 3D artists | Make a quick 3D preview from a single picture | A drawing of a prop becomes a rough 3D object
Product design | Designers | Check how an object looks from several sides | A product photo turned into a 3D splat
Research | ML researchers | Study distillation of flow-matching models | Compare a 5-step student with its 50-step teacher
Tool building | ComfyUI builders | Speed up existing TripoSplat workflows | Swap in the two LoRA files and set 5 steps

If you want the best quality and do not mind waiting, the original Idol at 50 steps is the card's quality reference, so you can skip this one.

Where to use it:

• ComfyUI, together with the Comfy-Org Idol files and its TripoSplat workflows
• Python, through the usage section of the model card
• Windows, macOS or Linux, with a GPU that suits your sampler

Use case | Who | What they get back
Fast image to 3D | 3D artists | A set of 3D Gaussians from one picture in about a second
Demo-style single objects | Designers | A rough 3D version of a single object
Training study | Researchers | A distilled model with its two phase experts

How to set it up: You need these first:

• ComfyUI installed on your computer
• The Idol files from the Comfy-Org Idol repack
• Python, for the Hugging Face download command

1. Place the Idol files from the Comfy-Org repack in their ComfyUI folders, as the table shows.
2. Download this repository with the hub tool.
3. Copy its two LoRA files from the loras folder into the ComfyUI loras folder.
4. Set the sampler to 5 steps and CFG 1.0.

pip install -U huggingface_hub
hf download speridlabs/idol-turbo

Title: Where each file goes inside the ComfyUI folder
File | Folder | Comes from
idol_fp16.safetensors | models/diffusion_models | Comfy-Org Idol repack
dino_v3_vit_h_bf16.safetensors | models/clip_vision | Comfy-Org Idol repack
flux2-vae.safetensors | models/vae | Comfy-Org Idol repack
idol_decoder_fp16.safetensors | models/vae | Comfy-Org Idol repack
idol_turbo_slow_comfy.safetensors | models/loras | This repository, loras folder
idol_turbo_fast_comfy.safetensors | models/loras | This repository, loras folder

How to use it: Run the workflow with 5 steps and CFG 1.0. Guidance is already built into the LoRA, so CFG 1.0 skips the extra guidance pass that would double the work. For best fidelity, split the run into two sampler stages, one per LoRA. For a simpler setup, use the slow LoRA alone.

Tool | What it does
idol_turbo_slow_comfy | LoRA for the high-noise steps, used at strength 1.0 in stage 1
idol_turbo_fast_comfy | LoRA for the low-noise steps, used in stage 2
KSampler (Advanced) | Splits the run into two stages, with add_noise on in stage 1 and off in stage 2

Example: You load one picture of a backpack into the Idol workflow with the slow LoRA and run 5 steps at CFG 1.0. What comes back is a 3D Gaussian version of the backpack in under a second on a fast GPU.

For newcomers and experts: Idol makes a 3D object from one picture, and the Turbo version does the same much faster. A good first step is the simple path, with only the slow LoRA, 5 steps and CFG 1.0.

Two rank-32 low-rank adapters, one for each noise phase, were trained to match the teacher's guided velocity, so guidance lives in the weights and the base model is untouched. Training used 29.4M parameters, 4,650 teacher records and 93 single-object images.

One picture of an object goes in
Five steps run with no guidance
The slow and fast LoRAs cover each phase
Up to a million 3D Gaussians come out

Good to know: Speed depends on the step count, and the card measured it on one H200.

Title: Seconds per object on one H200
4 steps: 0.66 s
5 steps: 0.75 s
8 steps: 1.0 s

• Evaluations: the base needs 100 network evaluations for 50 guided steps, and Turbo needs 5
• Strengths: single objects in the style of the demos, even kinds of object missing from training
• Weaknesses: a few results are softer than the 50-step teacher's, 4-step output may lean warm in color, and strongly saturated objects can look duller
• Untested split: the two-stage ComfyUI setup has not been run inside ComfyUI, so the slow-only path is the safe default
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the picture you give it and the model files in your ComfyUI folders

On Hugging Face, 13 November 2026

• Task: image to 3d
• Library: idol
• Downloads in the last 30 days: 565
• Likes: 57
• License: apache-2.0
• Based on: speridlabs/idol
• Tags: idol, gaussian-splatting, 3d, image-to-3d, turbo, distilled, comfyui, ml-intern

Links

• Model: https://huggingface.co/speridlabs/idol-turbo
• Files: https://huggingface.co/speridlabs/idol-turbo/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download speridlabs/idol-turbo

Then follow the usage in its model card. Nothing in this kit runs by itself.