Transformers Pull Request Adds Cellpose With SAM and DINOv3

What changed

Hugging Face Transformers is adding Cellpose support built on SAM and DINOv3 for cell segmentation. Work tracked as pull request 49624 opened in October 2026 under the title Add Cellpose (SAM + DINOv3). The goal is to bring a widely used microscopy segmentation stack into the same model hub patterns as language and vision models.

Cellpose helps labs outline cells in brightfield and fluorescence images. Combining segment anything models with DINOv3 features targets sharper masks on noisy lab plates. The PR is active review code, not yet a merged default release.

Issue and PR threads discuss API shape, weight loading, and test coverage expected before merge. Teams that already export Cellpose pipelines may still pin the classic package until Transformers ships a stable entry point.

Why it matters

Bio imaging teams want fewer one off scripts between training and deployment. A Transformers path could let shared MLOps pipelines serve segmentation beside other models if maintainers keep parity with the reference Cellpose behavior.

Regulated labs will still test on their own slide sets. A library merge does not replace sign off on cell counts used in assays.

Competition among open vision models is tight. DINOv3 features may help on more cell types, but labs must confirm on their own scopes and stains.

Who is affected

Microscopy core facility staff. Computational biologists who batch segment high content screens. ML engineers who standardize on Hugging Face hubs for model governance. Vendors who sell cloud analysis for cell health assays.

Grad students running Cellpose on laptops today may gain a simpler import path if docs stay beginner friendly.

What to do next

Follow PR 49624 for merge status and release tags. Do not retool production pipelines on a pre merge branch unless you accept churn.

If you depend on Cellpose today, keep your pinned version and export masks for regression tests. Compare new Transformers outputs on a fixed image panel when a release candidate lands.

What to watch

Watch the Transformers release note that mentions Cellpose after merge. Watch whether official weights land on the hub with clear license tags for commercial lab use.

Photo: National Cancer Institute / Pexels

Segmentation errors can skew dose response curves. Treat any new backend as a model change that triggers QA sign off in GxP settings.

Photo: Chokniti Khongchum / Pexels

Open a ticket with your IT team if outbound hub access is restricted. Weight downloads may need allow lists before pilots start.

Share benchmark images with maintainers only under your institution data rules. Public PR discussion is not a data sharing agreement.

SAM style models need GPU memory for large tiles. Core facilities should plan batch sizes before they move nightly jobs from laptops to shared servers.

If your assay counts cells per well, define accept limits when mask edges shift by a pixel. Small shifts can move totals on dense cultures.

Train new staff on both the classic Cellpose CLI and any new Transformers API once docs ship. Dual skill sets prevent downtime during migration windows.

Start with a small plate set that your lab already trusts. Run old and new masks side by side and mark diffs in a shared spreadsheet.

If you store PHI, keep images inside your VPN. Do not send patient slides to public cloud GPUs without a signed agreement.

Use simple names for mask layers in exports. Downstream stats scripts break when column names change between library versions.

Set a quarterly review to see if the PR merged. Slack reminders beat surprise upgrades during grant deadlines.

Keep a folder of gold standard masks for QA. Gold sets make regressions obvious when weights update on the hub.

Sources

• Primary — Hugging Face, Add Cellpose (SAM + DINOv3) (October 2026). Proposed Transformers integration for cell segmentation.
• Secondary — Hugging Face, Transformers repository (31 October 2026). Context for release process and testing norms.