DEXTUNE

One image in. A full 3D object out. 0.9 seconds. No photogrammetry rig. No multi-view capture. I fed a model a single PNG and it handed back 262,144 Gaussians before the progress bar finished drawing. It's Idol Turbo, and I got it running end-to-end on a Colab G4. Every number below is measured, not quoted from a README. What it is: @dextune distilled the 5-step version of speridlabs/idol (one image → 3D Gaussian splat). Two rank-32 LoRA experts, one for the high-noise phase and one for the low-noise phase, switching at t=0.5. That collapses ~100 network passes into 5. Apache-2.0. Measured on a G4 (RTX PRO 6000 Blackwell, 96 GB): 4 steps → 0.80 s 5 steps → 0.92 s 8 steps → 1.26 s 50 steps, CFG 3 (the un-distilled teacher) → 14.3 s 262,144 Gaussians → 17.8 MB .ply peak VRAM under 8 GB At 5 steps the time splits as: preprocess 0.14 s, DINOv3 encode 0.09 s, sampling 0.88 s, decode 0.10 s. So the second is almost entirely sampling, and it's 15.5× faster than the teacher while staying close on shape and color. The wall nobody mentions: the official SDK path refuses to load without a Hugging Face token that has accepted terms on two gated repos (facebook/dinov3-vith16plus and black-forest-labs/FLUX.2-dev). Anonymous downloads 401. So I rebuilt both encoders from ungated mirrors. DINOv3 from Comfy-Org's clip_vision tensors plus ONNX configs: 840.6 M params, 0 missing keys, 0 unexpected. FLUX.2 VAE from the Comfy single-file checkpoint with a hand-written remap to diffusers naming: 251/251 tensors, and a round-trip encode→decode MAE of 0.0105. No token, no gate, same weights. What it's good at: glossy vinyl characters with clean silhouettes and flat color. A baby dragon and an axolotl hold up from 360°. A chunky sneaker keeps its suede overlays, lace loops and lug pattern. Where it breaks: sharp metal edges and tiny high-frequency details. A rangefinder camera kept its dials but softened every edge. And it takes exactly one image — feed it a four-view sheet and it treats the whole sheet as one object. Crop to a single front or 3/4 view. Back views are invented, so they're the softest part of any turntable. A side-view input performs worse than front, because half the character isn't there and the model has to guess. The notebook has every step, the encoder rebuild and the VAE conversion included, and runs clean on a fresh G4. drive.google.com/file/d/1IgN… The real shift isn't 0.9 seconds. It's that generating a 3D asset now costs about the same as generating a sentence. I keep wondering what people build once a model is a throwaway.