v0.47.0

💪 Added

Depth Anything V2 🤝 inference

Depth estimation turns flat images into 3D information by calculating how far each pixel is from the camera. This adds an essential dimension to computer vision - distance.

Thanks to @reiffd7 we have Depth Anything V2 model in inference and Workflows ecosystem. Here is what the model brings into the table:

• ✨ Detailed results with sharp objects boundaries
• 💪 Robust performance across diverse scenes and lighting conditions
• ⚡ 10x faster processing than similar high-quality models
• 🪶 Lightweight (yet not available in Roboflow Hosted API, please use Roboflow Dedicated Deployment or deploy locally)

Why use it?

• Improve object detection with distance information
• Create 3D visualizations from single images
• Enable depth-aware applications (AR, robotics)
• Understand spatial relationships in scenes

Check out HF model card.

✨ New: OverlapFilter Block

@lou-roboflow added a new Overlap Filter block to the Workflows ecosystem.

The Overlap Filter removes objects that do not overlap a selected class. It is useful when you want to focus only on objects that are interacting with specific items.

How it works:

• You specify an overlap class (e.g., "bicycle").
• Only objects overlapping instances of that class are kept.
• The overlap class itself is removed from the results.

Example applications:

• Detecting people on bicycles (removing bicycles themselves).
• Identifying items on pallets.
• Finding passengers in cars, etc.

⚡ Speed-ups (with @codeflash-ai)

• ⚡️ Speed up function get_masks_intersection_up_to_dimension by 9% by @misrasaurabh1 in #1173
• ⚡️ Speed up function build_simple_operation by 68% by @misrasaurabh1 in #1171
• ⚡️ Speed up method OverlapManifest.describe_outputs by 536% in PR #1192 (OverlapBlockV1-2) by @codeflash-ai in #1193
• ⚡️ Speed up function get_average_bounding_box by 166% by @misrasaurabh1 in #1180
• optimize import time for inference by @KRRT7 in #1155

🧑‍🏭 Maintanence

• Preload hugginface IDs on inference startup by @grzegorz-roboflow in #1184
• Fix broken depth estimation import by @PawelPeczek-Roboflow in #1197
• Remove unused import from depth-estimation test module by @PawelPeczek-Roboflow in #1199
• Fix rf-detr to ignore predictions of background class by @Matvezy in #1178
• fixed depth inference integration tests, added endpoint for infer/dep… by @reiffd7 in #1201
• Add rf-detr base and large to weights upload docs by @capjamesg in #1200
• Disable depth estimation and bump version by @PawelPeczek-Roboflow in #1204
• Add CORRELATION_ID_LOG_KEY; fix typo for CORRELATION_ID_HEADER; control logging across modules with single env variable API_LOGGING_ENABLED by @grzegorz-roboflow in #1205
• Remove retired Claude version and add new ones by @brunopicinin in #1187
• Install codeflash optimization on CI by @misrasaurabh1 in #1105
• Feat/remove UUID validator from request correlation middleware for gcp and dd by @grzegorz-roboflow in #1183

🏅 New Contributors

• @codeflash-ai made their first contribution in #1193
• @lou-roboflow made their first contribution in #1192
• @KRRT7 made their first contribution in #1155

Full Changelog: v0.46.5...v0.47.0

v0.46.5

What's Changed

• Add docs about known issues with batch processing by @PawelPeczek-Roboflow in #1174
• fix url construction for get_roboflow_base_lora to be OS agnostic by @hansent in #1163
• Add Moondream2 by @capjamesg in #1146
• Assume no execution_time on error by @grzegorz-roboflow in #1150
• bump version 0.46.5 by @hansent in #1179

Full Changelog: v0.46.4...v0.46.5

v0.46.4

What's Changed

• drop openvino by @Matvezy in #1168
• version bump by @Matvezy in #1170

Full Changelog: v0.46.3...v0.46.4

v0.46.3

What's Changed

• Fix rfdetr postporccess by @Matvezy in #1159
• easy debug by @tonylampada in #1142
• ⚡️ Speed up function from_keypoints_detection_response by 7% by @misrasaurabh1 in #1154
• ⚡️ Speed up function convert_string_color_to_bgr_tuple by 30% by @misrasaurabh1 in #1153
• ⚡️ Speed up function count_specific_color_pixels by 60% by @misrasaurabh1 in #1152
• Bump version to 0.46.2 by @PawelPeczek-Roboflow in #1165
• Remove modules without any doscrtings from docs by @PawelPeczek-Roboflow in #1164
• Fix CPU builds which are failing for no reason :) by @PawelPeczek-Roboflow in #1166
• Loose wheel back by @PawelPeczek-Roboflow in #1167

Full Changelog: v0.46.1...v0.46.3

v0.46.1

What's Changed

• Add serialisation of timestamp in new UQL operations by @PawelPeczek-Roboflow in #1156
• Fix benchmark command to run with local images by @PawelPeczek-Roboflow in #1157

Full Changelog: v0.46.0...v0.46.1

v0.46.0

💪 Added

SmolVLM2 🤝 inference

SmolVLM2, a multimodal vision model developed by Hugging Face, is now available for use in Inference and Workflows. Thanks to its relatively compact size when compared to other multimodal models, SmolVLM2 is capable of running efficiently on less powerful devices, while still maintaining decent performance for a range of multimodal tasks, including VQA, document OCR, document VQA, and object counting.

You can use the model in Workflows:

Or you can run it directly with the inference Python package:

from PIL import Image
from inference.models.smolvlm.smolvlm import SmolVLM

smol_vlm = SmolVLM(api_key="API_KEY")
image = Image.open("dog.jpeg")
result = smol_vlm.predict(image, "How many dogs are in this image?")
print(result)

New operations in Workflows

Thanks to @grzegorz-roboflow contribution, Property Definition block was extended with new UQL operation - extracting times that elapsed since start of video from each video frame.

⚙️ Maintanence

• Cache results of 'inspect' by @grzegorz-roboflow in #1145
• Add Swagger for Batch Processing and Data Staging services by @PawelPeczek-Roboflow in #1147
• ⚡️ Speed up function collect_func_params by 10% by @misrasaurabh1 in #1128
• Fixing the address of jupyterlab in the Readme by @abelriboulot in #1130
• Make CACHE_METADATA_LOCK_TIMEOUT a parameter by @bigbitbus in #1135
• Feat/add execution duration to benchmark if avl by @bigbitbus in #1131
• Batch Processing docs update by @PawelPeczek-Roboflow in #1149
• Bump next from 15.2.3 to 15.2.4 in /inference/landing in the npm_and_yarn group across 1 directory by @dependabot in #1143

New Contributors

• @abelriboulot made their first contribution in #1130

Full Changelog: v0.45.3...v0.46.0

v0.45.3

What's Changed

• Cover collect_func_params with test by @grzegorz-roboflow in #1139
• Add env variable to block QWEN by @PawelPeczek-Roboflow in #1140

Full Changelog: v0.45.2...v0.45.3

v0.45.2

What's Changed

• Build CI to notify about changes in Google Colab env by @PawelPeczek-Roboflow in #1136
• Add gcloud setup to CI by @PawelPeczek-Roboflow in #1137
• Dependency issues in Colab 🔥 - second round by @PawelPeczek-Roboflow in #1133
• Adjust bitsandbytes to be installable on CPU by @PawelPeczek-Roboflow in #1138

Full Changelog: v0.45.1...v0.45.2

v0.45.1

What's Changed

• Cover center/bootom_left/bootom_right/top_left/top_right in property definition block by @grzegorz-roboflow in #1129
• Extend PropertyDefinition block to expose ExtractFrameMetadata operation by @grzegorz-roboflow in #1134

Full Changelog: v0.45.0...v0.45.1

v0.45.0

💪 Added

• Memory pressure safety valve by @bigbitbus in #1103

⚡️ Speed improvements by CodeFlash

• ⚡️ Speed up function extract_x_coordinate_of_detections_center by 35% by @misrasaurabh1 in #1114
• ⚡️ Speed up method Batch.remove_by_indices by 38% by @misrasaurabh1 in #1115
• ⚡️ Speed up method Batch.broadcast by 12% by @misrasaurabh1 in #1116

Other changes

• Added missing function call to load sky by @bigbitbus in #1113
• handle versionless model id through legacy routes by @grzegorz-roboflow in #1104
• Feat/pass service name received within request path param by @grzegorz-roboflow in #1117
• Fix usage collector test by @grzegorz-roboflow in #1118
• Fix issue with BC on OpenAPI spec in pydantic by @PawelPeczek-Roboflow in #1122
• Update RF-DETR docs by @capjamesg in #1121
• Do not pass default or empty source info by @grzegorz-roboflow in #1119

Full Changelog: v0.44.1...v0.45.0