v0.1.75 - Update Platform API contract (#96)

🌟 Summary

Release v0.1.75 brings video-frame skipping to cloud prediction, improves validation and saved video outputs, and synchronizes the Platform API contract with its Python SDK.

📊 Key Changes

• 🎞️ Cloud video prediction supports vid_stride: The option is sent to Platform, which predicts every Nth frame. Local output saving now uses the same frame selection, keeping predictions, labels, crops, and video frame rate aligned.
• 🛡️ Invalid prediction settings are rejected early: ul cloud predict checks arguments before uploading a local model or sending a Platform request. For example, vid_stride=0 fails before a prediction can incur a charge.
• 🔗 API and SDK updated to version 0.1.75: The OpenAPI contract and generated Python SDK now include optional vid_stride for model and deployment prediction requests; its minimum value is 1.
• 🧪 Expanded validation and clearer errors: CI now tests saved cloud-prediction outputs offline with CPU-based dependencies. CLI messages for authentication and connection failures are clearer, and SDK connection errors provide a fallback message when the underlying error is blank.

🎯 Purpose & Impact

• ✅ Makes cloud video predictions behave more consistently with local YOLO prediction, especially when saving results.
• 💸 Prevents invalid prediction arguments from triggering an unnecessary Platform request.
• 📦 Keeps the API contract, generated SDK, and package version in sync, so developers can rely on the SDK reflecting the supported API behavior.

What's Changed

• Honor vid_stride in ul cloud predict, test prediction saving, and clarify CLI errors by @glenn-jocher in #93
• Validate ul cloud predict arguments before the Platform request by @glenn-jocher in #94
• Send vid_stride to Platform predict and share the CLI test loopback server by @glenn-jocher in #95
• Update Platform API contract by @UltralyticsAssistant in #96

Full Changelog: v0.1.74...v0.1.75

v0.1.74 - Update Platform API contract (#92)

🌟 Summary

Release v0.1.74 expands the Platform API and Python SDK with Agent workflow management, adds dataset clustering metadata, and fixes saving prediction outputs from ul cloud predict.

📊 Key Changes

• 🤖 Agent workflow API and SDK: Added operations to list, save, and delete Agents—visual workflows combining blocks such as datasets, YOLO, LLMs, gates, and outputs. The Python SDK now supports these operations in both synchronous and asynchronous clients.
• 🧩 Dataset layout metadata: Added an optional cluster field to identify which cluster an image belongs to, including a value for scattered images. Older layouts may return null.
• 💾 More reliable cloud prediction outputs: ul cloud predict now uses public Results writers to save annotated images and videos, labels, and crops, and to display results. This avoids relying on internal inference-loop state that could cause crashes with newer Ultralytics versions.
• 🔄 Contract and SDK alignment: Updated the OpenAPI contract and Python SDK from 0.1.73 to 0.1.74, with CI checks for the generated SDK.

🎯 Purpose & Impact

• Developers can manage Agent workflows programmatically through the Platform API and Python SDK.
• Dataset clustering information is available to tools that consume dataset layouts.
• Cloud prediction results should save more reliably across supported output options, including video and text or crop exports.
• No model architecture or model-performance changes are included in this release.

What's Changed

• Save ul cloud predict outputs with the public Results writers by @JaviChulvi in #90
• Update Platform API contract by @UltralyticsAssistant in #92

Full Changelog: v0.1.73...v0.1.74

v0.1.73 - Update Platform API contract (#91)

🌟 Summary

v0.1.73 updates the Ultralytics Platform API and Python SDK with dataset-version comparison and more flexible exports, plus clearer CLI connection and authentication errors.

📊 Key Changes

• 🔍 Compare dataset versions: Added a dataset comparison endpoint and typed synchronous and asynchronous datasets.compare() methods. Responses can summarize added, removed, modified, or moved images, show label and dataset-header changes, and provide paginated results. A specific image can also be inspected across versions.
• 📦 Choose whether to prepare an export download: Dataset export creation now accepts an optional download setting. When downloads aren’t requested, the response may omit downloadUrl.
• 🔄 Expose restore progress context: Dataset response types now include restoreVersion, identifying the version an in-progress or interrupted restore applies.
• 🧭 Make CLI errors more actionable: A 401 response now includes guidance for creating and configuring an API key. Connection failures show the Platform URL and underlying error reason.
• ⚙️ Keep the SDK aligned with the API: The OpenAPI contract and regenerated Python SDK are versioned at 0.1.73, with CI validating the generated package.

🎯 Purpose & Impact

• Dataset comparisons make it easier to review what changed between saved versions and inspect how an individual image or its labels differ.
• Optional export downloads let users save a dataset version without also preparing a download, while keeping existing download behavior available by default.
• Clearer CLI messages help users diagnose invalid credentials and network or configuration problems more quickly.
• These are Platform API, SDK, and CLI improvements; the release does not introduce a model change.

What's Changed

• Add actionable 401 and connection error hints to the ul CLI by @JaviChulvi in #89
• Update Platform API contract by @UltralyticsAssistant in #91

Full Changelog: v0.1.72...v0.1.73

v0.1.72 - Update Platform API contract (#88)

🌟 Summary

Platform API and Python SDK v0.1.72 clarify deployment-renaming behavior, add a response field, and update a supported model ID.

📊 Key Changes

• 🚀 Deployment renaming is documented more clearly: Renaming changes the Platform URL slug, but preserves the Cloud Run service URL. The API now warns that existing Platform URL references—including those used by active workflows—will stop working.
• 🧩 Richer update responses: Deployment update responses may now include an optional deployment field with the deployment slug after renaming.
• ⚠️ Clearer conflict details: The API documentation now lists an existing deployment name as a possible cause of a conflict.
• 🤖 Updated model ID: Image prediction now lists gpt-6.1-sol instead of gpt-6-sol.
• 📦 SDK and contract version alignment: The OpenAPI contract and regenerated Python SDK move from 0.1.70 to 0.1.72.

🎯 Purpose & Impact

• 🔗 Users who rename deployments should update Platform URL references in integrations and active workflows; the underlying Cloud Run service URL remains unchanged.
• 🛠️ Python SDK users get updated types and API guidance, including the optional deployment field. The gpt-6.1-sol identifier is available in the SDK’s supported model list.
• ✅ Regenerating the SDK from the updated API contract and validating it in CI helps keep the published package consistent with the API.

What's Changed

• Update Platform API contract by @UltralyticsAssistant in #88

Full Changelog: v0.1.70...v0.1.72

v0.1.70 - Update Platform API contract (#87)

🌟 Summary

The v0.1.70 release synchronizes the Ultralytics Platform API contract and Python SDK with updated inference, prediction, similarity-search, and export details.

📊 Key Changes

• 💰 Added an optional estimated provider-cost field to inference responses.
• 🖼️ Updated image-prediction options: replaced claude-sonnet-5 with claude-sonnet-5-5, added grok-4.7, and raised the documented class limit from 100 to 200.
• 🔎 Clarified that similarity search can include public, personal, and team datasets, and better explained the similarity score.
• 📦 Updated export API documentation and generated SDK: removed the keras option and changed the example export format to TorchScript. The supported format list was reordered, not expanded or reduced.
• 🔄 Refreshed the generated Python SDK to version 0.1.70, matching the API contract. CI checks that the generated package stays aligned.
• 🧭 Updated contributor guidance and clarified CI, release, and contract-sync behavior.

🎯 Purpose & Impact

• Developers get SDK types and options that more accurately reflect the current Platform API, including cost information and the latest documented prediction models.
• Similarity-search documentation now makes the available dataset sources clearer.
• Export clients are less likely to rely on an option that is no longer in the contract; the format-list reorder itself does not change supported formats.
• Automated contract synchronization and validation help keep future SDK releases consistent with the API.

What's Changed

• Refresh AGENTS.md and align zh-CN CI triggers by @glenn-jocher in #85
• Update Platform API contract by @UltralyticsAssistant in #86
• Update Platform API contract by @UltralyticsAssistant in #87

Full Changelog: v0.1.65...v0.1.70

v0.1.65 - Update Platform API contract (#84)

🌟 Summary

Platform API contract 0.1.65 and its Python SDK were synchronized, with clearer API guidance and more helpful CLI errors.

📊 Key Changes

• Updated the API contract and regenerated the Python SDK to version 0.1.65, keeping the SDK aligned with the Platform API.
• Clarified dataset search and creation behavior: image search documentation now includes class names, and duplicate dataset or project slugs are documented as rejected with a 409 response. The obsolete require_exact_slug option was removed from dataset creation.
• Documented access requirements for training callbacks and checkpoint uploads, including workspace editor access where needed.
• Improved CLI error messages: invalid arguments now show valid argument names, and missing operations list the available choices.

🎯 Purpose & Impact

• Python SDK users get API definitions and documentation that more closely reflect current Platform behavior.
• Clearer slug-conflict and workspace-access guidance can help prevent failed requests and make errors easier to resolve.
• CLI users—and automated agents—can correct command mistakes faster without needing to open help separately.
• No model architecture or model-performance changes are included in this release.

What's Changed

• List valid arguments and operations in CLI errors by @JaviChulvi in #82
• Update Platform API contract by @UltralyticsAssistant in #84

Full Changelog: v0.1.62...v0.1.65

v0.1.62 - Update Platform API contract (#83)

🌟 Summary

Platform API v0.1.62 and its Python SDK add clearer dataset-import progress data and more image auto-annotation model options. ✨

📊 Key Changes

• Live dataset-import progress: Dataset responses can now include a processingProgress object with the current stage and percentage, plus optional processed, total, and cloud-object counts. It applies to imports of at least 10,000 images and is described as editor-only.
• More prediction model options: Image auto-annotation now lists SAM 3, SAM 3.1, Muse Spark 1.3, and MiMo v2.6 Pro/Flash among its supported choices.
• Clearer blur-preview behavior: API and SDK documentation now says blur previews provide refreshed signed URLs for each prepared image and thumbnail as they become ready.
• Synchronized SDK release: The Python SDK is updated to version 0.1.62 to match the API contract, with generated types and model options aligned. ✅

🎯 Purpose & Impact

• Dataset progress details can help users and tools better understand how large imports are moving along.
• Additional prediction options give users more choices for image auto-annotation.
• The synchronized API contract and SDK make these updates easier to discover and use from Python.

What's Changed

• Update Platform API contract by @UltralyticsAssistant in #83

Full Changelog: v0.1.57...v0.1.62

v0.1.57 - Pass auto-annotate request as body in live test (#81)

🌟 Summary

v0.1.57 adds face-blurring tools for datasets and makes cloud training easier to start, monitor, and download. It also updates the SDK and fixes a live test for the new dataset-processing request format.

📊 Key Changes

• 🧪 CI compatibility fix: Updated the live test to pass dataset batch requests in the new body format, keeping its missing-model check working after the API change.
• 🙂 Dataset face blurring: The Platform API now supports blurring faces across a dataset or a selected image. Users can preview up to six images before applying blur and adjust detection confidence and blur-box size.
• ☁️ More flexible cloud training: ul cloud train now starts training and returns with the model URI, run link, selected GPU, estimated cost, and a command for downloading results later. Add watch to follow epoch progress and download results when training finishes.
• 📥 Download results later: The new ul cloud download command can wait for an existing training run and save its best checkpoint and available training results locally.
• 🔄 Updated Platform SDK and API contract: The Python SDK now supports the expanded dataset workflows and revised batch request format. Similar-image searches can also prepare an image embedding when needed.

🎯 Purpose & Impact

• ✅ Users can start cloud training without staying connected until it completes, then retrieve results when convenient.
• 🔒 Face blurring provides a way to prepare dataset images with added privacy protection, while previews help users review images before applying changes.
• 🧩 The regenerated SDK reflects the latest API capabilities; callers of dataset batch processing should use the new body request format.
• 🛠️ The CI fix ensures live tests continue to validate the updated API contract.

What's Changed

• Detach ul cloud train by default and add watch and ul cloud download by @JaviChulvi in #77
• Update Platform API contract by @UltralyticsAssistant in #79
• Update Platform API contract by @UltralyticsAssistant in #80
• Fix live CI after predict/batch body became a union by @glenn-jocher in #81

Full Changelog: v0.1.54...v0.1.57

v0.1.54 - Update Platform API contract (#78)

🌟 Summary

Expanded Platform image prediction and batch-annotation model support, while making the Python SDK versioning and regeneration process more reliable and consistent. 🚀

📊 Key Changes

• Added hosted prediction models 🧠
  Image prediction and dataset batch annotation now recognize additional model identifiers, including:

  • florence2
  • owlv2
  • yoloe26x
  • groundingdino
  • Several additional hosted generative and vision models

• Improved prediction documentation 📚
  The API and generated Python SDK now explain that hosted models:

  • Use model-specific confidence thresholds
  • Return detections without confidence scores
  • May return partial results when generative output is truncated
  • Do not support depth datasets for annotation conversion

• Regenerated the Python SDK 🐍
  The SDK’s type hints and documentation now expose the expanded model_id choices for both synchronous and asynchronous image prediction and dataset batch-annotation methods.

• Updated the Platform API contract to 0.1.54 🔄
  The ultralytics-platform package now matches the upstream API contract version exactly.

• Strengthened CI validation ✅
  CI now verifies that:

  • The API contract uses a valid MAJOR.MINOR.PATCH version
  • The SDK package version matches the API contract
  • No independent SDK version override is configured
  • Generated SDK files remain synchronized with the contract

• Clarified release and recovery procedures 🛠️
  Documentation now establishes the upstream Platform API contract as the single source of truth for SDK releases. SDK-only or generator changes require a coordinated API contract version bump rather than an independent SDK patch increment.

🎯 Purpose & Impact

• More annotation options: Users can choose from a broader range of hosted vision models for image prediction and automated dataset labeling. 🎯
• Better Python developer experience: Updated type hints and documentation make supported model identifiers easier to discover and use.
• More dependable releases: Enforcing a single API and SDK version prevents mismatches between the deployed Platform API and the published Python package.
• Safer automation: Deterministic SDK regeneration and stricter CI checks reduce generated-code drift and help ensure releases contain the intended API behavior.
• Important compatibility note: This release primarily expands Platform API and SDK capabilities; it does not introduce a new core YOLO model architecture. Users should confirm that their selected hosted model is available for their Platform workflow and account.

What's Changed

• Use the upstream API contract as the only SDK release version by @glenn-jocher in #75
• Update Platform API contract by @UltralyticsAssistant in #78

Full Changelog: v0.1.52...v0.1.54

v0.1.52 - Update Platform API contract (#76)

🌟 Summary

Updated the Ultralytics Platform API contract and Python SDK to v0.1.52, improving version consistency, reproducibility, and release validation. 🚀

📊 Key Changes

• 📦 Python SDK version bumped from 0.1.51 to 0.1.52.
• 📄 OpenAPI contract version updated from 0.1.49 to 0.1.52.
• ⚙️ Removed the hardcoded generator version from openapi.config.json, allowing the release process to manage SDK versions automatically.
• 🔐 Updated the OpenAPI SHA-256 checksum to verify the integrity of the refreshed contract.
• ✅ The SDK is deterministically regenerated and validated by CI before release, ensuring the published package matches the approved API definition.

🎯 Purpose & Impact

• 🔄 Keeps the Platform API specification and Python SDK aligned under the same release version.
• 🛡️ Improves confidence that the generated SDK accurately reflects the current API contract.
• 🤖 Enables automatic patch-version increments for SDK-only updates while preserving versioning consistency for API changes.
• 👩‍💻 Users can upgrade to ultralytics-platform 0.1.52 for the latest contract metadata and a more reliable integration with the Ultralytics Platform.

What's Changed

• Update Platform API contract by @UltralyticsAssistant in #76

Full Changelog: v0.1.51...v0.1.52