{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny","slug":"shi-labs--oneformer_ade20k_swin_tiny","name":"oneformer_ade20k_swin_tiny","type":"model","url":"https://huggingface.co/shi-labs/oneformer_ade20k_swin_tiny","page_url":"https://unfragile.ai/shi-labs--oneformer_ade20k_swin_tiny","categories":["image-generation"],"tags":["transformers","pytorch","oneformer","vision","image-segmentation","dataset:scene_parse_150","arxiv:2211.06220","license:mit","endpoints_compatible","deploy:azure","region:us"],"pricing":{"model":"open_source","free":true,"starting_price":null},"status":"active","verified":false},"capabilities":[{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_0","uri":"capability://image.visual.unified.image.segmentation.with.task.conditioning","name":"unified-image-segmentation-with-task-conditioning","description":"Performs semantic, instance, and panoptic segmentation on images using a single unified transformer-based architecture that conditions on task-specific prompts. The model uses a Swin Transformer backbone (tiny variant) with a OneFormer decoder that processes image features through cross-attention mechanisms guided by task embeddings, enabling a single model to handle multiple segmentation tasks without task-specific fine-tuning or separate model checkpoints.","intents":["I need to segment scenes into semantic categories (e.g., sky, building, tree) without training separate models for each task","I want to detect and delineate individual object instances in an image while also understanding scene semantics","I need panoptic segmentation that combines stuff (background) and thing (object) classes in a unified output","I want to run segmentation inference on edge devices or resource-constrained environments with a lightweight model"],"best_for":["computer vision researchers prototyping multi-task segmentation pipelines","teams building scene understanding systems for robotics or autonomous systems","developers deploying segmentation models to edge devices or mobile platforms","organizations working with ADE20K dataset or similar indoor scene datasets"],"limitations":["Swin Tiny backbone limits receptive field and feature resolution compared to larger variants (Swin Base/Large), reducing accuracy on small objects or fine details","Model trained exclusively on ADE20K indoor scenes; performance degrades significantly on outdoor or domain-shifted images","No built-in support for real-time inference optimization (quantization, pruning, or TensorRT conversion) — requires external tooling","Requires full image as input; does not support region-of-interest or patch-based inference for memory efficiency","Task conditioning via text embeddings adds ~50-100ms latency per inference compared to task-specific models"],"requires":["PyTorch 1.9+","transformers library 4.25+","CUDA 11.0+ for GPU inference (CPU inference supported but slow)","Minimum 4GB VRAM for batch size 1 inference","PIL/Pillow for image loading and preprocessing"],"input_types":["image (RGB, 8-bit, variable resolution)","task token/embedding (semantic, instance, or panoptic)","optional: image metadata (resolution, aspect ratio)"],"output_types":["segmentation mask (H×W integer tensor with class indices)","instance IDs (H×W tensor for instance segmentation)","class probabilities (H×W×num_classes float tensor)","panoptic segmentation map (H×W with encoded stuff/thing IDs)"],"categories":["image-visual","vision-transformers"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_1","uri":"capability://image.visual.ade20k.scene.parsing.with.150.class.taxonomy","name":"ade20k-scene-parsing-with-150-class-taxonomy","description":"Segments images into 150 semantic classes from the ADE20K dataset taxonomy, including fine-grained scene categories (e.g., 'kitchen', 'bedroom', 'bathroom') and object classes (e.g., 'chair', 'table', 'window'). The model maps pixel-level features to this 150-class space through a learned classification head trained on ADE20K's densely annotated indoor scene images, enabling detailed scene understanding for indoor environments.","intents":["I need to parse indoor scenes into fine-grained semantic categories for scene understanding applications","I want to identify specific room types and furniture in images for smart home or robotics applications","I need to extract scene context (e.g., 'this is a kitchen with a table and chairs') for downstream reasoning tasks","I want to benchmark my segmentation model against the ADE20K standard evaluation protocol"],"best_for":["indoor robotics teams building scene understanding for navigation and manipulation","smart home developers analyzing room layouts and object placement","researchers evaluating segmentation models on the ADE20K benchmark","teams building scene graph or visual relationship detection systems"],"limitations":["Taxonomy is fixed to ADE20K's 150 classes; no support for custom class vocabularies or fine-tuning on new domains without retraining","Performance drops significantly on outdoor scenes, non-English labeled images, or domains not represented in ADE20K training data","Class imbalance in ADE20K (e.g., rare classes like 'escalator' or 'fountain') leads to poor recall on underrepresented categories","No hierarchical class relationships exposed; treats all 150 classes as independent, missing semantic relationships (e.g., 'chair' is furniture)","Requires preprocessing to ADE20K's expected input distribution; images with different lighting, color spaces, or aspect ratios may degrade accuracy"],"requires":["ADE20K class label mapping (150 classes)","Image preprocessing: resize to 512×512 or 1024×1024, normalize to ImageNet statistics","PyTorch 1.9+","transformers 4.25+"],"input_types":["RGB image (variable resolution, typically 512×512 to 2048×2048)"],"output_types":["segmentation mask (H×W integer tensor, values 0-149 for ADE20K classes)","class confidence scores (H×W×150 float tensor)","class name strings (mapped from integer indices)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_2","uri":"capability://image.visual.lightweight.swin.tiny.backbone.inference","name":"lightweight-swin-tiny-backbone-inference","description":"Executes image feature extraction using a Swin Transformer Tiny backbone (28M parameters) with hierarchical window-based self-attention, enabling efficient inference on resource-constrained devices. The backbone processes images through 4 stages with shifted window attention patterns, reducing computational complexity from O(n²) to O(n log n) compared to dense attention, while maintaining spatial locality through local window operations.","intents":["I need to run segmentation inference on edge devices, mobile phones, or embedded systems with limited compute","I want to reduce model size and latency for real-time segmentation applications (e.g., video processing)","I need to deploy the model to cloud inference endpoints with cost constraints (e.g., AWS Lambda, Azure Functions)","I want to understand the trade-off between model size, inference speed, and accuracy for my application"],"best_for":["edge ML engineers deploying models to Jetson, Raspberry Pi, or mobile devices","teams building real-time video segmentation pipelines with latency constraints","cost-conscious teams deploying to serverless inference platforms","researchers studying efficient vision transformer architectures"],"limitations":["Swin Tiny has 40% fewer parameters than Swin Base, resulting in ~5-8% absolute mIoU drop on ADE20K (48.5% vs 53%+)","Window-based attention limits global receptive field; small objects or fine details may be missed compared to dense attention models","Inference latency on CPU is 2-3 seconds per image; GPU required for sub-500ms latency","Memory footprint is ~450MB for model weights + ~2GB for intermediate activations during inference; not suitable for <1GB RAM devices","No quantization or pruning applied; INT8 quantization requires external tools and may degrade accuracy by 2-3%"],"requires":["PyTorch 1.9+","transformers 4.25+","Minimum 2GB RAM for inference","CUDA 11.0+ for GPU acceleration (optional but recommended)"],"input_types":["image tensor (3×H×W, float32, normalized to ImageNet statistics)"],"output_types":["feature pyramid (4 levels: 1/4, 1/8, 1/16, 1/32 resolution)","hierarchical embeddings (dimensions: 96, 192, 384, 768 for each stage)"],"categories":["image-visual","automation-workflow"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_3","uri":"capability://image.visual.multi.scale.feature.aggregation.with.decoder","name":"multi-scale-feature-aggregation-with-decoder","description":"Aggregates multi-scale features from the Swin Tiny backbone through a OneFormer decoder that fuses features across 4 hierarchical levels using cross-attention and self-attention mechanisms. The decoder progressively upsamples features while attending to task-specific embeddings, enabling the model to combine low-level details with high-level semantic context for accurate segmentation at original image resolution.","intents":["I need to combine features from multiple scales (1/4, 1/8, 1/16, 1/32) to preserve both fine details and semantic context","I want to condition feature fusion on task-specific information (semantic vs instance vs panoptic) without separate decoders","I need to upsample low-resolution feature maps back to original image resolution while maintaining spatial accuracy","I want to understand how multi-scale features contribute to segmentation accuracy for different object sizes"],"best_for":["vision researchers studying multi-scale feature fusion architectures","teams building segmentation models that need to handle objects at varying scales","developers optimizing segmentation accuracy for small objects or fine boundaries","engineers analyzing feature importance and attention patterns in segmentation models"],"limitations":["Cross-attention between task embeddings and features adds ~100-150ms latency per inference compared to simple concatenation-based fusion","Decoder requires storing intermediate feature maps at all 4 scales, increasing peak memory usage by ~1.5GB during inference","No explicit mechanism to handle scale imbalance; small objects may still be underrepresented if not explicitly weighted in loss function","Upsampling from 1/32 to 1× resolution through 5 stages introduces cumulative interpolation artifacts; bilinear upsampling may blur fine boundaries","Task conditioning via embeddings is not interpretable; difficult to debug why certain tasks perform better than others"],"requires":["PyTorch 1.9+","transformers 4.25+","Multi-scale feature maps from Swin Tiny backbone (4 levels)","Task embedding vectors (e.g., from text encoder or learned embeddings)"],"input_types":["feature pyramid from backbone (4 levels: 96, 192, 384, 768 channels)","task embedding (e.g., 256-dim vector for semantic/instance/panoptic)","original image resolution metadata"],"output_types":["upsampled feature maps (1× resolution, 256-512 channels)","per-pixel class logits (H×W×num_classes)","attention maps (optional, for interpretability)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_4","uri":"capability://image.visual.batch.image.segmentation.with.variable.resolution","name":"batch-image-segmentation-with-variable-resolution","description":"Processes multiple images of varying resolutions in a single batch through dynamic padding and batching logic, enabling efficient throughput for inference pipelines. The model handles images with different aspect ratios by padding to a common size within each batch, then crops predictions back to original dimensions, avoiding the need to process each image individually.","intents":["I need to segment a batch of images with different resolutions efficiently without processing them one-by-one","I want to maximize GPU utilization by batching images of varying sizes","I need to process images from different sources (e.g., webcam, file system, API) with inconsistent dimensions","I want to benchmark throughput (images/second) for my segmentation pipeline"],"best_for":["teams building batch inference pipelines for image processing services","developers optimizing throughput for video frame segmentation","data annotation platforms processing user-uploaded images of varying sizes","researchers benchmarking segmentation model efficiency"],"limitations":["Padding to common batch size introduces wasted computation on padded regions; batch efficiency decreases as aspect ratio variance increases","Maximum batch size limited by GPU VRAM; typical batch size is 4-8 on 8GB GPUs, 16-32 on 24GB+ GPUs","No built-in support for dynamic batching based on image size; requires manual bucketing of images by resolution","Padding artifacts may appear at image boundaries if not handled carefully; requires post-processing to remove padding effects","Throughput gains from batching plateau after batch size ~16 due to memory bandwidth limitations"],"requires":["PyTorch 1.9+","transformers 4.25+","GPU with sufficient VRAM (8GB minimum for batch size 4, 24GB+ for batch size 16+)","Image preprocessing pipeline (resize, pad, normalize)"],"input_types":["batch of RGB images (variable H×W, 3 channels, 8-bit)","batch size (integer, 1-32 typical)"],"output_types":["batch of segmentation masks (B×H×W integer tensors)","batch of class probabilities (B×H×W×150 float tensors)"],"categories":["image-visual","automation-workflow"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_5","uri":"capability://image.visual.instance.segmentation.with.panoptic.decoding","name":"instance-segmentation-with-panoptic-decoding","description":"Generates instance-level segmentation masks by decoding per-pixel class predictions and instance IDs, enabling distinction between individual object instances of the same class. The model produces both semantic segmentation (class per pixel) and instance IDs, which are combined to create panoptic segmentation that unifies stuff (background) and thing (object) classes with unique instance identifiers.","intents":["I need to count and locate individual objects in an image (e.g., 'how many chairs are in this room?')","I want to extract bounding boxes or masks for each object instance for downstream processing","I need panoptic segmentation that combines semantic and instance information in a single output","I want to track object instances across video frames for temporal consistency"],"best_for":["robotics teams building object detection and manipulation pipelines","video analysis platforms tracking objects across frames","scene understanding systems that need both semantic and instance information","annotation tools generating instance masks for dataset creation"],"limitations":["Instance segmentation accuracy depends on class-specific instance separation; overlapping or touching objects may be merged into single instances","No explicit instance tracking across frames; temporal consistency requires external tracking algorithms (e.g., Hungarian matching)","Instance IDs are not stable across different inference runs; cannot be used for cross-image instance matching","Performance on small instances (<50 pixels) is poor due to Swin Tiny's limited receptive field; requires larger backbone variants","Panoptic quality metric (PQ) is lower than specialized instance segmentation models (e.g., Mask R-CNN) due to unified architecture trade-offs"],"requires":["PyTorch 1.9+","transformers 4.25+","Post-processing logic to convert class predictions + instance IDs to panoptic masks","Optional: connected components analysis for instance ID assignment"],"input_types":["RGB image (variable resolution)","task token set to 'panoptic' or 'instance'"],"output_types":["panoptic segmentation map (H×W integer tensor with encoded stuff/thing IDs)","instance ID map (H×W integer tensor, unique ID per instance)","semantic class map (H×W integer tensor, 0-149 for ADE20K classes)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_6","uri":"capability://image.visual.task.conditioned.inference.with.text.prompts","name":"task-conditioned-inference-with-text-prompts","description":"Conditions model behavior on task-specific text prompts (e.g., 'semantic segmentation', 'instance segmentation', 'panoptic segmentation') by encoding prompts into embeddings and using them to modulate attention in the decoder. This enables a single model checkpoint to perform multiple segmentation tasks without task-specific fine-tuning, with task selection happening at inference time through prompt selection.","intents":["I want to switch between semantic, instance, and panoptic segmentation at inference time without reloading models","I need to adapt the model to new tasks by providing task-specific prompts without retraining","I want to understand how task conditioning affects model predictions and attention patterns","I need to build interactive applications where users can select segmentation tasks dynamically"],"best_for":["interactive computer vision applications with dynamic task selection","multi-task learning researchers studying task conditioning mechanisms","teams building flexible segmentation APIs that support multiple task modes","developers exploring prompt-based model adaptation without fine-tuning"],"limitations":["Task conditioning via text embeddings requires a text encoder (e.g., CLIP, BERT); adds ~50-100ms latency and 100-200MB model size","Limited to predefined task prompts; cannot generalize to arbitrary new tasks without retraining","Task embeddings are not interpretable; difficult to understand what aspects of the embedding drive task-specific behavior","Performance gap between task-conditioned model and task-specific models is 2-5% mIoU; unified architecture trades accuracy for flexibility","No mechanism to combine multiple tasks (e.g., 'semantic + instance segmentation'); tasks are mutually exclusive"],"requires":["PyTorch 1.9+","transformers 4.25+","Text encoder (e.g., CLIP text encoder, BERT) for prompt embedding","Predefined task prompt vocabulary (semantic, instance, panoptic)"],"input_types":["RGB image (variable resolution)","task prompt string (e.g., 'semantic segmentation', 'instance segmentation')","optional: task embedding vector (if pre-computed)"],"output_types":["task-specific segmentation output (semantic mask, instance mask, or panoptic map)","task embedding used for conditioning (for interpretability)"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_7","uri":"capability://tool.use.integration.huggingface.model.hub.integration.with.pretrained.weights","name":"huggingface-model-hub-integration-with-pretrained-weights","description":"Provides seamless integration with Hugging Face Model Hub, enabling one-line model loading with pretrained weights via the transformers library. The model is hosted on Hugging Face with full model card documentation, inference examples, and community discussions, allowing developers to load and use the model without manual weight downloading or configuration.","intents":["I want to load a pretrained segmentation model with a single line of code without downloading weights manually","I need to access model documentation, training details, and benchmark results from a centralized hub","I want to use the model with Hugging Face inference APIs (e.g., Hugging Face Inference Endpoints) without local setup","I need to version-control my model usage and track which model checkpoint I'm using in my application"],"best_for":["developers building quick prototypes with pretrained models","teams using Hugging Face ecosystem tools (transformers, datasets, accelerate)","researchers sharing models and enabling reproducibility","organizations deploying models via Hugging Face Inference Endpoints"],"limitations":["Requires internet connection to download model weights on first use; no offline mode without pre-caching","Model weights are cached in ~/.cache/huggingface/hub; requires ~450MB disk space for Swin Tiny variant","Hugging Face API rate limits may apply for high-volume inference requests; not suitable for production without dedicated endpoints","Model card documentation is community-maintained; may contain outdated or incomplete information","No version pinning mechanism; model updates on Hub may break backward compatibility if not carefully managed"],"requires":["transformers library 4.25+","PyTorch 1.9+","Internet connection for initial model download","Hugging Face account (optional, for private model access)"],"input_types":["model identifier string (e.g., 'shi-labs/oneformer_ade20k_swin_tiny')"],"output_types":["loaded model object (transformers.PreTrainedModel)","model configuration (transformers.PretrainedConfig)","tokenizer/processor (if applicable)"],"categories":["tool-use-integration","automation-workflow"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_8","uri":"capability://automation.workflow.pytorch.and.onnx.export.for.deployment","name":"pytorch-and-onnx-export-for-deployment","description":"Supports export to PyTorch and ONNX formats for deployment across different inference frameworks and hardware platforms. The model can be exported to ONNX for inference on CPU, mobile, or specialized hardware (e.g., NVIDIA TensorRT, CoreML for iOS), enabling deployment flexibility beyond PyTorch-only environments.","intents":["I need to deploy the model to production environments that don't support PyTorch (e.g., C++ servers, mobile apps)","I want to optimize inference latency using ONNX Runtime or TensorRT on specific hardware","I need to run the model on mobile devices (iOS, Android) or embedded systems","I want to quantize or prune the model for deployment on resource-constrained devices"],"best_for":["production ML engineers deploying models to diverse hardware platforms","mobile app developers integrating segmentation into iOS/Android applications","teams optimizing inference latency using specialized inference engines","organizations deploying models to edge devices or embedded systems"],"limitations":["ONNX export requires careful handling of dynamic shapes; variable-resolution inputs may not export cleanly without fixed input dimensions","ONNX Runtime performance varies by hardware; may not match PyTorch performance on all platforms without optimization","Quantization (INT8) during export typically reduces accuracy by 2-5%; requires careful calibration and validation","ONNX export is not officially supported by Hugging Face for all model architectures; may require custom export scripts","Mobile deployment (iOS/Android) requires additional conversion steps (ONNX → CoreML/TFLite) and model size reduction techniques"],"requires":["PyTorch 1.9+","onnx library for export","onnxruntime for inference (optional, for testing)","TensorRT (optional, for NVIDIA GPU optimization)","CoreML tools (optional, for iOS deployment)"],"input_types":["PyTorch model (transformers.PreTrainedModel)","example input tensor (for tracing/scripting)"],"output_types":["ONNX model file (.onnx)","PyTorch scripted model (.pt)","model metadata (input/output shapes, opset version)"],"categories":["automation-workflow","tool-use-integration"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"hf-model-shi-labs--oneformer_ade20k_swin_tiny__cap_9","uri":"capability://automation.workflow.azure.endpoints.compatible.inference.deployment","name":"azure-endpoints-compatible-inference-deployment","description":"Compatible with Azure Machine Learning endpoints for serverless inference deployment, enabling integration with Azure's managed inference infrastructure. The model can be deployed to Azure ML endpoints with automatic scaling, monitoring, and integration with Azure's authentication and logging systems.","intents":["I want to deploy the model to Azure ML endpoints for serverless inference without managing infrastructure","I need automatic scaling based on inference request volume","I want to integrate the model with Azure's monitoring, logging, and authentication systems","I need to expose the model as a REST API endpoint for downstream applications"],"best_for":["teams using Azure cloud infrastructure for ML workloads","organizations requiring managed inference with automatic scaling","developers building production ML services with minimal DevOps overhead","enterprises needing integration with Azure's security and compliance systems"],"limitations":["Azure ML endpoint costs scale with compute hours; may be expensive for high-volume inference without careful resource planning","Cold start latency for serverless endpoints is 5-10 seconds; not suitable for real-time applications requiring <1 second response time","Requires Azure account and familiarity with Azure ML SDK; steeper learning curve than local deployment","Model size (450MB) may exceed Azure Functions limits; requires Azure ML endpoints or container-based deployment","No built-in support for batch inference optimization; requires custom batching logic or Azure Batch integration"],"requires":["Azure subscription with ML workspace","Azure ML SDK (azureml-sdk)","Docker for containerization (optional, for custom environments)","Model registered in Azure ML model registry"],"input_types":["image (base64-encoded or URL)","task parameter (semantic/instance/panoptic)"],"output_types":["segmentation mask (base64-encoded or URL)","JSON response with predictions and metadata"],"categories":["automation-workflow","tool-use-integration"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":45,"verified":false,"data_access_risk":"high","permissions":["PyTorch 1.9+","transformers library 4.25+","CUDA 11.0+ for GPU inference (CPU inference supported but slow)","Minimum 4GB VRAM for batch size 1 inference","PIL/Pillow for image loading and preprocessing","ADE20K class label mapping (150 classes)","Image preprocessing: resize to 512×512 or 1024×1024, normalize to ImageNet statistics","transformers 4.25+","Minimum 2GB RAM for inference","CUDA 11.0+ for GPU acceleration (optional but recommended)"],"failure_modes":["Swin Tiny backbone limits receptive field and feature resolution compared to larger variants (Swin Base/Large), reducing accuracy on small objects or fine details","Model trained exclusively on ADE20K indoor scenes; performance degrades significantly on outdoor or domain-shifted images","No built-in support for real-time inference optimization (quantization, pruning, or TensorRT conversion) — requires external tooling","Requires full image as input; does not support region-of-interest or patch-based inference for memory efficiency","Task conditioning via text embeddings adds ~50-100ms latency per inference compared to task-specific models","Taxonomy is fixed to ADE20K's 150 classes; no support for custom class vocabularies or fine-tuning on new domains without retraining","Performance drops significantly on outdoor scenes, non-English labeled images, or domains not represented in ADE20K training data","Class imbalance in ADE20K (e.g., rare classes like 'escalator' or 'fountain') leads to poor recall on underrepresented categories","No hierarchical class relationships exposed; treats all 150 classes as independent, missing semantic relationships (e.g., 'chair' is furniture)","Requires preprocessing to ADE20K's expected input distribution; images with different lighting, color spaces, or aspect ratios may degrade accuracy","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.5698063124444298,"quality":0.45,"ecosystem":0.5000000000000001,"match_graph":0.25,"freshness":0.75,"weights":{"adoption":0.35,"quality":0.2,"ecosystem":0.1,"match_graph":0.3,"freshness":0.05}},"observed_outcomes":{"matches":0,"success_rate":0,"avg_confidence":0,"top_intents":[],"last_matched_at":null},"maintenance":{"status":"active","updated_at":"2026-05-24T12:16:22.766Z","last_scraped_at":"2026-05-03T14:23:00.161Z","last_commit":null},"community":{"stars":null,"forks":null,"weekly_downloads":null,"model_downloads":248429,"model_likes":17}},"distribution":{"claim_url":"https://unfragile.ai/submit?claim=shi-labs--oneformer_ade20k_swin_tiny","compare_url":"https://unfragile.ai/compare?artifact=shi-labs--oneformer_ade20k_swin_tiny"}},"signature":"d6Jj/FU9wK49LkIVGSposAsYQxu8KKPy60MIYwPjumExIxYJ4di3WnzcuFSwUa4kZYCR8UnYWg1OwvcM0AuqDg==","signedAt":"2026-06-20T02:24:58.811Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/shi-labs--oneformer_ade20k_swin_tiny","artifact":"https://unfragile.ai/shi-labs--oneformer_ade20k_swin_tiny","verify":"https://unfragile.ai/api/v1/verify?slug=shi-labs--oneformer_ade20k_swin_tiny","publicKey":"https://unfragile.ai/api/v1/trust-passport-public-key","spec":"https://unfragile.ai/trust","schema":"https://unfragile.ai/schema.json","docs":"https://unfragile.ai/docs"}}