Idol Turbo distills image-to-3D to 5 steps: ~0.75s for up to 1M Gaussians on H200
_akhaliq · x · 2026-11-04
Idol Turbo, a distilled few-step version of speridlabs/idol, is now on Hugging Face (akhaliq/idol-turbo, Apache-2.0).
- One object photo in, up to a million 3D Gaussians out — 5 guidance-free steps at 0.75s on an H200 (0.66s at 4 steps, 1.0s at 8)
- Cuts network evaluations from 100 (50 steps × CFG) to 5, a 20× NFE reduction
- Uses a Glance-style phase-LoRA recipe (arXiv 2512.02899): two rank-32 experts (Slow-LoRA for high noise, Fast-LoRA for low noise, switching at t=0.5) trained by regression against the teacher's CFG-composed velocity; only 29.4M trainable params, base model untouched
- The v5 recipe trained on 4,650 teacher trajectories across 93 diverse single-object images for 2,000 steps (25 min on one H200), fixing the earlier version's poor generalization from a 10-image training set
- Verified compatible with the Comfy-Org/Idol repack and TripoSplat ComfyUI workflows at strength 1.0
Related event: IDOL Turbo distills single-image 3D generation to 0.75 seconds(2 posts)→
More from Multimodal
- Quantized Qwen-Image-2.1-Turbo (GGUF/FP8/MLX) Trends on Hugging Face — abenzerps · 2026-11-04
- 132-Page Tutorial on Generative Models for Music Audio Lands on Hugging Face — amaai-lab · 2026-11-04
- Filmmaker uses Claude Opus to direct AI movie: still needs tons of reshoots for 150 minutes — MosskeepForest · 2026-11-04
- Open-source TaoMate-FVSR restores face videos with audio conditioning and visual dubbing — Apprehensive_Sky892 · 2026-11-04
- AI-subagent QA misses a LoRA training image with a hilariously reversed butt — illumnat · 2026-11-04
- The famous 'Clark Kent unfolds into Superman' AI clip gets a reverse: Mr. Bean scrunches — NathanpmYoung · 2026-11-04