Idol Turbo β€” structural-admitted Edgework package

One image in, one 3D Gaussian splat out: the Speri Labs Idol pipeline with the Idol Turbo phase-LoRA distill (5 guidance-free steps), packaged for the Forkjoin/Edgework mesh with its admission receipt.

Honest status: STRUCTURAL admission only. This package ships two artifacts: the loadable safetensors set, and a KNOT v1 container (idol-turbo.knot, 3.64 GB, 2,138 tensors, round-trip verified) that is an archive β€” the sovereign runtime has no Idol pipeline, so the knot is not executable. No visible-tier admission (no 3DGS renderer or recognisability oracle), and the shipped sample is a 1-step, cfg-0 probe render β€” not a quality reference.

What is admitted, and what is not

Claim State Evidence
artifact is a well-formed 3DGS PLY ADMITTED (structural) monster-guard qspec verify-modality idol-3d.qspec.json armchair.ply β†’ monster-guard.modality-receipt/v1
finite values, zero normals, unit quaternions, distinct xyz, bounded log-scale/opacity ADMITTED receipt checks block (see receipt.json)
recognisable object / prompt-faithful / visible render NOT CLAIMED β€” refused by construction verifier refuses gate_tier: visible; no renderer exists yet
knot container PROVIDED (archive only) idol-turbo.knot β€” 3,638,843,041 B, 2,138 tensors, KNOT v1, round-trip verified (knot-receipt.json); no sovereign Idol loader exists, so it is not executable

Files

File Bytes SHA-256
config.json 1,335 ddd89dcc6dcf4a5ac8d951b8c59afa1eceafe377615c5912c59f4e1f5ae35b32
denoiser.safetensors 1,482,258,308 ded2cacf671bf3978ebc62729087abdbd318ef4752e6805b4c9fc8ffa51fe716
decoder.safetensors 203,171,696 9da4d7676bfcadb1c3fdfbff796e6257dadf82d26ec05be7717c79b2b13a1c4e
decoder_gs.safetensors 949,083,036 3087eb8ee5763206a53e4078acf2a1bf382dab5147d407ec13a2cf39c04cc544
birefnet.safetensors 444,466,856 44bbc97b041f879975d0d6d538b8e66a25eb04d1f49aab40697544cdd97eaa8e
idol_turbo_lora.safetensors 117,484,528 2e4b089211c5d79354f8b67d512f9b436248ac072cea171f56d393a248f7bf66
turbo.py 5,817 d607930134bc1375b5b9d652a2bab8ca0b38f7eceef973686a5dba96b6b2d264
idol-turbo.knot (archive) 3,638,843,041 c5a19cb7ef8d2f18521331c0aa67d7dadf6e0debbb0aa06ee8417e46b51e3210

Also included: idol-3d.qspec.json (the modality gate), receipt.json (the checked monster-guard.modality-receipt/v1), probe_splat.py (standalone structural probe) and sample/armchair.ply (the admitted artifact).

Structural receipt (abridged)

{
  "schema": "monster-guard.modality-receipt/v1",
  "modality": "image-to-3d",
  "gate_tier": "structural",
  "artifact": { "sha256": "4175d235d6fa31cfd568397f543f74c1bf86854902060c3ad83ea37a40c9c231", "byte_length": 2228639 },
  "checks": {
    "format": "binary_little_endian", "gaussians": 32768, "bytes_per_gaussian": 68,
    "finite": true, "zero_normals": true,
    "rotation_norm": { "min": 0.9999998299829844, "max": 1.000000154735748 },
    "scale": { "min": -7.013115882873535, "max": -2.1643764972686768 },
    "opacity": { "min": -13.815509796142578, "max": 13.80231761932373 },
    "extent": { "min": -0.5290173888206482, "max": 0.5134075880050659 },
    "distinct_xyz": 32768
  },
  "ok": true
}

Measured producer run

Device: 1Γ— Xeon E5530 (2009, no AVX), 8 threads, 31 GB RAM β€” CPU-only, no GPU. Pipeline: idol-turbo weights, DINOv3 ViT-H+/16 via the public timm/vit_huge_plus_patch16_dinov3.lvd1689m port, FLUX.2 VAE via kaiyuyue/FLUX.2-dev-vae, fp32 with the SDK's hardcoded self.dtype = torch.float16 overridden.

Measurement Value
one denoise step (24 blocks, T=8192) 328.5 s (37.7 GFLOPS effective)
end-to-end, 1 step, 32,768 Gaussians 636.3 s
PLY 2,228,639 B = 68 B/Gaussian + 415 B header
predicted turbo (5 steps), same box ~32.5 min

Intended serving path (CF-only)

A standard-4 Cloudflare Container runs the render as isolated requests under the 900 s ceiling: prepare (encoders + seeded noise) β†’ step Γ— N (one denoiser forward each, latent checkpointed β€” 512 KiB fp32) β†’ decode (octree + Gaussian decoders β†’ PLY). Skymesh/edgework serves the artifact from R2 behind a Protocol69 FOIL cache key. Workers never execute the model.

Usage

# upstream SDK β€” requires access to facebook/dinov3-vith16plus and black-forest-labs/FLUX.2-dev
from idol import IdolPipeline
pipeline = IdolPipeline.from_pretrained("<this-repo>")
output = pipeline("photo.jpg", mode="real", steps=5, cfg_strength=0.0)
output.gaussian.save_ply("photo.ply")

cfg_strength=0.0 with the turbo LoRA is the guidance-free 5-step setting. The step-isolated runner (prepare/step/decode) lives in the Edgework repository at apps/aeon-puppet-idol-cf/runner.

Limitations

  • No visible-tier admission; do not present the sample as a render.
  • DINOv3 and FLUX.2 VAE come from public mirrors, not the gated Meta/BFL repos; encoder parity against the gated originals is unverified.
  • The sample artifact is 1 step at cfg=0; it exists to prove the structural gate and the artifact path, not image quality.
  • The .knot is a container, not a loadable model: KNOT v1 holds the 2,138 fp32 tensors with per-tensor shape/digest metadata, but no sovereign Idol pipeline can execute it yet. Loading it as a model would be a claim the runtime cannot honour.
  • License and credit: base weights and turbo distill are Apache-2.0 from Speri Labs (idol, idol-turbo). This package adds packaging, the qspec and the structural receipt only.
Downloads last month
3
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for forkjoin-ai/idol-turbo

Base model

speridlabs/idol
Finetuned
(8)
this model