Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Hyper3D / Deemos / CLAY

CLAY Large-base

CLAY Large-base is a named geometry diffusion model evaluated in the CLAY paper. It is a base model trained on the full processed 3D dataset.

3D generation
restrictedClosed weights3D generation
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
8 of 23 specification fields collected
3D output formatsNot collectedAwaiting source review
Text input limitNot collectedAwaiting source review
Input billingNot collectedHyper3D-associated research publication
Output billingNot collectedNo inferred price

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
Not collected
Maximum input tokens
Not collected
Maximum output
Not collected
Reasoning
Not collected
Tool calling
Not collected
Structured output
Not collected
Streaming
Not collected
Batch processing
Not collected
Fine-tuning
Not collected
Knowledge cutoff
Not collected
Release date
Not collected

Publication or observed application version does not establish an original weight release date.

Parameters
853,000,000Source for Parameters (opens in a new tab)

DiT size from Table 1; excludes the separate 82M geometry VAE and material model.

Architecture
24-layer latent diffusion transformer, width 1536, 16 attention heads; feed-forward expansion 4×. Geometry uses a separate multiresolution VAE.Source for Architecture (opens in a new tab)
Languages
Not collected
Context caching
Not collected
Usage metering
Not collected

No price or currently callable endpoint for this exact research edition verified. Current Rodin API rates are not inherited.

Data retention
Not collected

No hosting, retention or enterprise contract verified for this research edition.

Input modalities
TextSource for Input modalities (opens in a new tab)

Text-to-shape evaluation model. Adapter support is not inherited from XL-P.

Output modalities
3D geometrySource for Output modalities (opens in a new tab)
Training sequence length
1024 geometric latent positions in Table 3 evaluation; 64 channels per position.Source for Training sequence length (opens in a new tab)

Geometric representation, not prompt-token capacity.

Published training data
527,000 filtered ShapeNet/Objaverse objects; remeshed and annotated through the paper’s processing pipeline.Source for Published training data (opens in a new tab)
Geometry resolution
Paper inference: 100 denoising steps, occupancy sampled on a 512³ grid, then Marching Cubes.Source for Geometry resolution (opens in a new tab)

Reported research configuration; not a hosted service limit.

PBR materials
Not supportedSource for PBR materials (opens in a new tab)

This record is the geometry DiT. CLAY Material Diffusion supplies PBR textures separately.

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Hyper3D-associated research publication

Provider source (opens in a new tab)
API model ID
Not collected

No callable selector or immutable download for this research edition verified.

Access
Research paper/project page. No currently callable endpoint or downloadable weights verified for this exact edition.
Regions
Processing region not established
Endpoint availability
Research edition with no verified public checkpoint download or native API selector.Source for Endpoint availability (opens in a new tab)

Endpoint pricing has not been collected. Check the provider for applicable rates.

Data & enterprise terms

Research access is not assigned the commercial Rodin API retention policy.

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

Hyper3D-associated research publication

Not verified

No model-specific prompt-cache contract or price established.

How it works
Not collected

No model-specific prompt-cache contract or price established.

Minimum reusable input
Not collected

No model-specific prompt-cache contract or price established.

Lifetime & refresh
Not collected

No model-specific prompt-cache contract or price established.

How to enable it
Not collected

No model-specific prompt-cache contract or price established.

Usage fields
Not collected

No model-specific prompt-cache contract or price established.

Billing rules
Not collected

No model-specific prompt-cache contract or price established.

Limits & data boundaries
Not collected

No model-specific prompt-cache contract or price established.

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 14 Oct 2026

What the measurements say

CLAY paper / Render-FID

CLAY paper authors · author-reported
6.5732score

Table 3, Large-base, 1024 geometric latent positions, 16K text-shape validation pairs. Lower is better. Eight rendered views for render metrics; point-cloud feature metrics use PointNet++. Paper-specific protocol, not an independently reproduced or cross-provider leaderboard.

Evaluation source (opens in a new tab)

CLAY paper / Render-KID

CLAY paper authors · author-reported
2.3617×10³

Table 3, Large-base, 1024 geometric latent positions, 16K text-shape validation pairs. Lower is better. Eight rendered views for render metrics; point-cloud feature metrics use PointNet++. Paper-specific protocol, not an independently reproduced or cross-provider leaderboard.

Evaluation source (opens in a new tab)

CLAY paper / P-FID

CLAY paper authors · author-reported
0.865score

Table 3, Large-base, 1024 geometric latent positions, 16K text-shape validation pairs. Lower is better. Eight rendered views for render metrics; point-cloud feature metrics use PointNet++. Paper-specific protocol, not an independently reproduced or cross-provider leaderboard.

Evaluation source (opens in a new tab)

CLAY paper / P-KID

CLAY paper authors · author-reported
1.6377×10³

Table 3, Large-base, 1024 geometric latent positions, 16K text-shape validation pairs. Lower is better. Eight rendered views for render metrics; point-cloud feature metrics use PointNet++. Paper-specific protocol, not an independently reproduced or cross-provider leaderboard.

Evaluation source (opens in a new tab)

CLAY paper / CLIP(I-T)

CLAY paper authors · author-reported
0.2358similarity

Table 3, Large-base, 1024 geometric latent positions, 16K text-shape validation pairs. Higher is better. Eight rendered views for render metrics; point-cloud feature metrics use PointNet++. Paper-specific protocol, not an independently reproduced or cross-provider leaderboard.

Evaluation source (opens in a new tab)

CLAY paper / ULIP-T

CLAY paper authors · author-reported
0.1559similarity

Table 3, Large-base, 1024 geometric latent positions, 16K text-shape validation pairs. Higher is better. Eight rendered views for render metrics; point-cloud feature metrics use PointNet++. Paper-specific protocol, not an independently reproduced or cross-provider leaderboard.

Evaluation source (opens in a new tab)

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Qualified unreleased-weight classification as reviewed: the official OpenCLAY tree contains only README and example images; the linked organization model listing is empty. The paper documents trained models. This does not assert a permanent proprietary license or rule out future releases; public research publication and integration code are not weight releases.

Classification evidence (opens in a new tab)

Lifecycle dates

Research paper first submitted30 May 2024Source (opens in a new tab)

Sources reviewed