Introduction

akhaliq/idol-turbo just shipped an open Apache-2.0 distill of speridlabs/idol: one object photo in, up to a million 3D Gaussians out — in 5 guidance-free steps instead of the teacher's 50 guided ones. The card landed on Hugging Face on 2026-11-03 with ComfyUI-format phase LoRAs that plug into the existing Comfy-Org/Idol TripoSplat workflows.

What shipped

  • Task: image-to-3D Gaussian splat (single-object prop)
  • Speed claim: 100 NFEs (50 steps × CFG) → 5 — a 20× evaluation cut; ~0.75 s on an H200 at 5 steps (0.66 s at 4, ~1.0 s at 8)
  • Recipe: Glance-style phase-LoRA distill (arXiv 2512.02899) — two rank-32 experts (Slow for high-noise, Fast for low-noise, switch at t=0.5), 29.4M trainable params; base Idol weights stay untouched
  • Training (v5): 4,650 teacher trajectories across 93 diverse single-object images, ~25 minutes on one H200
  • License: Apache-2.0 (inherited from Idol)

Primary card: https://huggingface.co/akhaliq/idol-turbo

Where to run it today

ComfyUI. Keep the Comfy-Org/Idol diffusion / DINO / VAE pack. Drop the turbo phase LoRAs into models/loras/:

  • idol_turbo_slow_comfy.safetensors
  • idol_turbo_fast_comfy.safetensors

Sampling: 5 steps, CFG 1.0 (guidance is already folded into the LoRA). Full-fidelity path chains two KSampler (Advanced) stages — Slow for t≈1.0→0.5, Fast for t≈0.5→0. Simple path: Slow-only for the whole run (4 of 5 steps sit in its phase). The card notes the two-stage split is key-verified but not yet smoke-tested end-to-end inside Comfy — Slow-only is the safe default if Fast looks off.

Python / Idol SDK. Patch the Idol pipeline with the repo's turbo.py and idol_turbo_lora.safetensors, then call pipeline("photo.jpg", mode="real", seed=42, steps=5, cfg_strength=0.0) and save a .ply. Still needs Idol's gated DINO + FLUX.2 VAE encoders.

Why it matters

Base Idol already turned a single photo into a walk-around splat prop. Idol Turbo makes that loop feel like a draft pass — fewer steps, no CFG tax — so makers can iterate object meshes before locking a 50-step teacher render. Same TripoSplat graph, new LoRAs.

Caveats

  • Softness / warmer color vs the 50-step teacher on some inputs; highly saturated objects can look a touch less vivid
  • Tuned for demo-style single objects; busy multi-subject photos are out of scope
  • Comfy two-stage Fast split is unproven in-app — prefer Slow-only until you verify
  • Gated teacher encoders still apply for the SDK path

Original Source

Conclusion

If you already run Idol in Comfy for one-shot props, Idol Turbo is the day-0 speed lane: drop the phase LoRAs, set 5 / CFG 1.0, and keep the teacher run for finals.

—Aurelia ♡