SeamFlow

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

API task JSON is not caller-defined structured model output.

Streaming

Not collected

SSE reports task progress; streamed geometry output is not established.

Batch processing

Not collected

Multi-view inputs and queued tasks do not establish a discounted batch API.

Fine-tuning

Not collected

No customer model-weight fine-tuning contract documented.

Knowledge cutoff

Not collected

Release date

Not collected

Original release date not established by a later API or research announcement.

Parameters

Approximately 712 millionSource for Parameters (opens in a new tab)

Full model count reported in Appendix A.

Architecture

Flow matching on mesh-edge probabilities. A 28-block DiT uses width 1024, 16 attention heads and a 4096-wide SwiGLU feed-forward layer; local edge tokens attend to global shape features.Source for Architecture (opens in a new tab)

Languages

Not collected

Context caching

Not collected

No documented cache controls or cache price.

Output modalities

Mesh seam probabilities, UV layoutSource for Output modalities (opens in a new tab)

Input modalities

3D meshSource for Input modalities (opens in a new tab)

Research controls

50-step Euler sampling; alternate noise seeds produce different seam layouts. Local seam inpainting can hold unmasked areas fixed.Source for Research controls (opens in a new tab)

Research procedures, not hosted API request fields.

Training dataset

350,000 filtered Objaverse meshes with UVs and curated seam layouts; 2,000 held-out test meshes.Source for Training dataset (opens in a new tab)

Published training data

Training curation excludes meshes over 50,000 faces, degenerate geometry, self-intersections and implausible UV layouts.Source for Published training data (opens in a new tab)

Dataset curation threshold, not an API input limit or guaranteed maximum supported mesh.

Training hardware

8 NVIDIA H100 GPUsSource for Training hardware (opens in a new tab)

Training hardware, not inference deployment requirements.

Research output capabilities

Predicted edge probabilities are thresholded at 0.5. The evaluation pipeline uses SLIM for parameterization and Blender for UV packing.Source for Research output capabilities (opens in a new tab)

Endpoint availability

Research publication; project page marks code and models as forthcoming. No callable SeamFlow API selector is verified.Source for Endpoint availability (opens in a new tab)

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

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

No benchmark observations have been added for this model. We do not turn a vendor description into a performance score.

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.

Why it is included

The authors’ project page explicitly marks code and models as forthcoming. This record describes currently unreleased weights, not a permanently proprietary license. Recheck when the promised release appears. The project-page template license does not license model weights.

Lifecycle dates

Sources reviewed