Idol timing, 31 Oct 2026: what ran, and on what.

Model       speridlabs/idol, revision 2731c2291fc65a797c5cb627a5e5526fc1490902 (Apache-2.0, Hugging Face, image to 3D)
Code        speridlabs/idol-sdk, commit 4702a7f00850e34b6e83a2170cc123c3e9570744 (29 Oct 2026), no local changes
            idol/pipeline.py sha256 091fc0fe7d8f66f5f9f63795a91f9b654e12cf879d94bf03624f39c360ddd4a1
            idol/sample.py   sha256 4c32003164917631b5b650640ee7ef44d2dfd0198cd70539f6d02d6d7a56fd62
Encoders    facebook/dinov3-vith16plus-pretrain-lvd1689m, snapshot c807c9eeea853df70aec4069e6f56b28ddc82acc (gated on Hugging Face, bf16 as the SDK default)
            black-forest-labs/FLUX.2-dev, used for its VAE (gated on Hugging Face, the revision was not recorded)
            Both machines had the same Idol and DINOv3 snapshots (same revision hashes). The runs used HF_HUB_OFFLINE=1.
Software    Python 3.13.5, torch 2.14.1+cu130, triton 3.8.0, transformers 5.22.1, diffusers 0.42.1 (the same on both machines)
Attention   None of flash-attn, xformers, sageattention or kernels is installed. The SDK uses PyTorch scaled_dot_product_attention
            (its README: same result as with flash-attn, slower).
            The SDK's fast sampling is on (the default): torch.compile, batched guidance, cuDNN attention and TF32 matmuls in the denoiser.
Cards       NVIDIA GeForce RTX 5060 Ti 16 GB, driver 595.71.05
            NVIDIA GeForce RTX 5090 32 GB, driver 610.43.02
Start       GPU memory in use before the first run: 4 MiB on both cards (idle)

Settings    mode synthetic (the SDK default for computer generated pictures), steps 50, cfg_strength 3.0, rescale_t 3.0, seed 42,
            262,144 Gaussians (SDK default) for every picture, then 1,048,576 (SDK maximum) for a subset (8 pictures in run 1, every second one, and all 6 in run 2).
            Everything else is the SDK default. The same script and the same settings ran on both cards.
Timing      time.time() around the call pipe(image, ...), with torch.cuda.synchronize() before and after. Saving the PLY file is after the timer.
            The first picture of a process is timed on its own because torch.compile runs then. The median leaves it out.
Memory      nvidia-smi memory.used of the card, polled every 0.2 s while the pictures run (includes the CUDA context), and
            torch.cuda.max_memory_allocated() as the second number.
Clips       Turntables are rendered from the 262,144-Gaussian PLY files with scripts/render_turntable.py.txt (a small PyTorch splat renderer, no
            CUDA extension, not the author's viewer): 240 frames, one turn, 720 x 720, 30 fps, 8 s, then H.264 with ffmpeg.