difference in the result between inference API and preview in hub

[vipin-prabhakaran]

Search before asking

• I have searched the HUB issues and discussions and found no similar questions.

Question

i have a Ultralytics Dedicated Inference API which i started with an end point tokyo ,japan with yolov11,i am passing images which is getting generated from base 64 to this to get results of detection ,i am facing with an issue , i have used the same script i got it from hub portal so all configurations are same

when i am making request with an image generated from base64 i am getting different confidence levels like for PPE detection from api i am getting a confidence value .45 for a class but when i take out the same image and put it in hub preview for my model its giving .85 for the same class ,i am not getting the issue becoz i used same image ,i am using python scrypt which will be available for the api to make the request

becoz of this client is losing confidence on my model and saying values are wrong from api

can any one help me to figure out what will be the issue

Additional

No response

[UltralyticsAssistant]

👋 Hello @vipin-prabhakaran, thank you for raising an issue about Ultralytics HUB 🚀! This is an automated response to help you get started, and an Ultralytics engineer will follow up shortly.

Please visit our HUB Docs to learn more:

• Quickstart. Start training and deploying YOLO models with HUB in seconds.
• Datasets: Preparing and Uploading. Learn how to prepare and upload your datasets to HUB in YOLO format.
• Projects: Creating and Managing. Group your models into projects for improved organization.
• Models: Training and Exporting. Train YOLOv5 and YOLOv8 models on your custom datasets and export them to various formats for deployment.
• Integrations. Explore different integration options for your trained models, such as TensorFlow, ONNX, OpenVINO, CoreML, and PaddlePaddle.
• Ultralytics HUB App. Learn about the Ultralytics App for iOS and Android, which allows you to run models directly on your mobile device.
  • iOS. Learn about YOLO CoreML models accelerated on Apple's Neural Engine on iPhones and iPads.
  • Android. Explore TFLite acceleration on mobile devices.
• Inference API. Understand how to use the Inference API for running your trained models in the cloud to generate predictions.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us investigate differences between HUB preview and the Dedicated Inference API. See the steps to reproduce. Specifically, please include:

• Exact HUB model you’re previewing (model name/ID) and confirmation it’s the same checkpoint used by the API (e.g., YOLO11)
• Exact inference parameters used in both places: confidence threshold, IoU/NMS settings, image size, classes filter, augmentation, etc.
• The single test image used for both runs (or its base64 source before encoding/decoding)
• Your minimal API request example (Python or cURL) showing endpoint region (Tokyo), request headers, payload, and any preprocessing (base64 encoding steps)
• The raw API response for that image, plus a HUB preview screenshot showing the differing confidence
• Timestamps of the two runs so we can check server logs if needed

If this is a ❓ Question, please provide as much information as possible, including dataset, model, and environment details so that we can provide the most helpful response.

We try to respond to all issues as promptly as possible. Thank you for your patience! 🙏💡

[pderrenger]

Hi @vipin-prabhakaran — thanks for the clear write-up, and sorry for the confusion. When the same model and identical image bytes are used with the same inference settings, the HUB preview and the Dedicated Inference API should return essentially identical results. A jump like 0.45 → 0.85 usually indicates a mismatch in one of these areas:

Checklist to align preview vs API

• Use the exact same image bytes

  • If you’re starting from base64, be sure you’re sending the decoded raw bytes as the multipart file, not a re-encoded/quality-changed image. Even a JPEG re-save can shift confidences a lot.

  • Quick example for posting decoded base64 bytes:

    import io, base64, requests

    endpoint = "YOUR_DEDICATED_ENDPOINT_URL" # from Deploy tab
    headers = {"x-api-key": "API_KEY"}
    data = {
    "model": "https://hub.ultralytics.com/models/MODEL_ID",
    "imgsz": 640,
    "conf": 0.25,
    "iou": 0.45,
    }

    img_bytes = base64.b64decode(B64_STRING) # decoded raw bytes
    files = {"file": ("image.jpg", io.BytesIO(img_bytes))}
    r = requests.post(endpoint, headers=headers, data=data, files=files)
    print(r.json())

  • Tip: Compare the “shape” returned in the API JSON with the HUB preview image’s dimensions. If they differ, the inputs are not byte-identical.

• Match inference parameters exactly
  • Explicitly set imgsz, conf, and iou in your API call to match preview. The API defaults are imgsz=640, conf=0.25, iou=0.45. Different imgsz or IoU can change which box survives NMS and its confidence.
  • See the Inference API arguments in the docs for reference: Ultralytics HUB Inference API.
• Confirm the same model artifact
  • Ensure the Deploy tab and your request’s model URL both point to the same MODEL_ID and checkpoint. If you recently retrained or switched checkpoints, restart the Dedicated endpoint (Stop, then Start) to pick up changes.
• Control test
  • Send the same image to the Shared Inference API (https://predict.ultralytics.com) with the same params and compare results to your Dedicated endpoint. This helps isolate any endpoint-specific variance. See the Ultralytics HUB Inference API guide for quick cURL/Python examples.

Why confidence can differ

• Different image bytes (re-encoding, resized, color-space, EXIF orientation handling) or different imgsz/IoU can result in a different candidate box surviving NMS, hence a different confidence. Confidence threshold (conf) only filters outputs; it doesn’t change the predicted confidence value itself, but NMS IoU can change which box is returned.

Could you share the following so we can reproduce and pinpoint this quickly?

• Your HUB model link or MODEL_ID.
• The Dedicated endpoint URL (region is fine; redact any sensitive parts).
• The exact Python/cURL request you’re sending (redact API key).
• One example image: the base64 you send to the API and the same image file you upload to preview.
• The full JSON response from the API for that image.

References

• How to call the API and control imgsz/conf/iou: See the Ultralytics HUB Inference API docs.
• General tips on aligning deployment behavior: See Best Practices for Model Deployment.

Once we have those details, I’ll run the same image through both paths and report back with the exact cause and fix. We’ll get your client’s confidence back on track quickly. 🙌

[vipin-prabhakaran]

hi @pderrenger

thanks for your response ,i will try to pass base 64 directly with out changing it to image ,meanwhile i can give my modelid ijercziuepidijkvdfwp and i am using tokyo-japan regeon ,and i am using the exact python script which is given under the deploy tab with no difference like this url = "https://predict-mymodelid-regeon-an.a.run.app"
headers = {"x-api-key": "mykey"}
data = {"imgsz": 640, "conf": 0.25, "iou": 0.45}
with open("/path/original_1756216244 _1.jpg", "rb") as f:
response = requests.post(url, headers=headers, data=data, files={"file": f})
response.raise_for_status(). if u need more info please let us know we tried with base 64 way which u told,which also giving the same rsults as we tried with image like {
"images": [
{
"results": [
{
"box": {
"x1": 1036.96741,
"x2": 1437.06848,
"y1": 3374.2085,
"y2": 3869.92163
},
"class": 11,
"confidence": 0.86338,
"name": "shoes"
},
{
"box": {
"x1": 892.39923,
"x2": 2198.60034,
"y1": 876.43341,
"y2": 2445.1106
},
"class": 8,
"confidence": 0.85652,
"name": "no_jacket"
},
{
"box": {
"x1": 1610.04834,
"x2": 1990.70776,
"y1": 3412.60449,
"y2": 3894.94263
},
"class": 11,
"confidence": 0.72222,
"name": "shoes"
},
{
"box": {
"x1": 1969.20862,
"x2": 2370.39868,
"y1": 1725.73059,
"y2": 2560.4436
},
"class": 0,
"confidence": 0.52635,
"name": "gloves"
},
{
"box": {
"x1": 797.4892,
"x2": 1091.83032,
"y1": 1773.52185,
"y2": 2561.15625
},
"class": 0,
"confidence": 0.41786,
"name": "gloves"
}
],
"shape": [
4000,
3000
],
"speed": {
"inference": 1798.32673,
"postprocess": 1.36912,
"preprocess": 2.95892
}
}
],
"metadata": {
"functionTimeAlive": 1142.499,
"functionTimeCall": 2.067,
"imageCount": 1,
"model": "https://hub.ultralytics.com/models/mymodel",
"version": {
"python": "3.11.10",
"torch": "2.8.0+cpu",
"torchvision": "0.23.0+cpu",
"ultralytics": "8.3.183"
}
}
}. but the same image i am getting like this kind of results in my preview

[pderrenger]

Hi @vipin-prabhakaran — thanks for the extra detail and the sample JSON, that helps a lot.

Good news first: your Dedicated Inference endpoint is clearly receiving the original 4000×3000 bytes and running cleanly (ultralytics 8.3.183, CPU). The confidences you pasted (e.g., shoes 0.86, no_jacket 0.86, gloves 0.72/0.53/0.42) look reasonable. A large jump like 0.45 → 0.85 for the same object typically comes from one of the following, so here’s how to lock everything down:

1. Ensure both paths use the same checkpoint

• In Ultralytics HUB, open your model’s Deploy tab.
• Set Selected checkpoint to Best (recommended) or the exact checkpoint you want.
• Click Stop Endpoint, then Start Endpoint to reload that exact artifact for your Dedicated API.
• Preview uses the selected checkpoint too, so this keeps both sides in sync.

2. Include the explicit model parameter on requests

• Although Dedicated endpoints are bound to a single model, passing the model explicitly removes any ambiguity:
  data = {
  "model": "https://hub.ultralytics.com/models/ijercziuepidijkvdfwp",
  "imgsz": 640,
  "conf": 0.25,
  "iou": 0.45
  }

3. Match preview sliders with API params

• Align imgsz, conf, and iou exactly with what the Preview shows (conf and iou control filtering/NMS; they don’t change the raw confidence produced by the network, but they can change which box survives NMS, which can look like a “confidence change” if a different candidate box is selected).
• If your image has small or dense objects, try imgsz=1280 in both Preview and API and compare.

4. Quick A/B control test (Shared vs Dedicated)

• Send the same bytes to both endpoints to isolate any endpoint-specific variance:

  import io, requests

  file_path = "/path/original_1756216244_1.jpg"
  with open(file_path, "rb") as f:
  img_bytes = f.read()

  headers = {"x-api-key": "API_KEY"}
  data = {
  "model": "https://hub.ultralytics.com/models/ijercziuepidijkvdfwp",
  "imgsz": 640,
  "conf": 0.25,
  "iou": 0.45
  }

  Shared

  r_shared = requests.post(
  "https://predict.ultralytics.com",
  headers=headers, data=data, files={"file": ("image.jpg", io.BytesIO(img_bytes))}
  )

  Dedicated

  r_ded = requests.post(
  "YOUR_DEDICATED_ENDPOINT_URL",
  headers=headers, data=data, files={"file": ("image.jpg", io.BytesIO(img_bytes))}
  )

  print("Shared:", r_shared.json())
  print("Dedicated:", r_ded.json())

5. If a delta persists

• Please share:
  • A screenshot of Deploy > Selected checkpoint.
  • The exact Preview settings (imgsz/conf/iou).
  • The API JSON you pasted is great; if possible, also capture the Preview’s JSON for the same image (or a full-screen screenshot where the confidence number is visible on the same box).
  • The specific detection you’re comparing (class name and rough box location), so we can verify we’re looking at the same candidate box on both sides.

For reference on request/response fields and available args, see the Ultralytics HUB Inference API docs. For tips on aligning deployment behavior and why NMS/resize/encoding can affect which box is returned, see Best Practices for Model Deployment.

If you set the same checkpoint and params, the Preview and Dedicated API should match closely. Share the above details and I’ll help pinpoint anything still off. We’ll get this squared away for your client. 🙌

[vipin-prabhakaran]

hi @pderrenger thanks for your response ,i will try it out and update on the same ,i just want to have a clarity on one thing ,are you resizing the images to 640 while uploading to preview ,to the API request we are not sending resized images ,our data set is trained for 640 px sizes ,do we need to resize the images also to 640 before sending to the api to match the result becoz i am seing this is the area where there is mismatch if the images are getting resized on preview,i really not seing any settings on the deploy to set the check point ,i really didnt get that ,if indeed we need to resize the image will it be 640px 640 px to match the results . any help will be greatly appreciated thanks in advance

[pderrenger]

Hi @vipin-prabhakaran — great questions, thanks for the details.

Short answer on resizing

• You do not need to resize images client-side. Both HUB Preview and the Dedicated Inference API internally resize/letterbox the input to imgsz (default 640) before running inference, then scale boxes back to the original image and return the original shape in the JSON. So sending your 4000×3000 image is fine as long as imgsz matches between Preview and API. See the Ultralytics HUB Inference API docs for the available arguments and defaults.
• If you prefer to pre-resize yourself, you must replicate Ultralytics’ letterbox behavior (preserve aspect ratio + pad). A naive 4000×3000 → 640×640 warp can hurt accuracy and won’t match Preview.

How to align Preview and API

• In Preview, set the sliders to the exact values you send to the API: imgsz (e.g., 640 or 1280), conf (e.g., 0.25), and iou (e.g., 0.45). Different imgsz or IoU can change which box survives NMS, which can look like a “confidence change.”

• In your API call, include the model parameter explicitly to remove any ambiguity:

  data = {
  "model": "https://hub.ultralytics.com/models/ijercziuepidijkvdfwp",
  "imgsz": 640,
  "conf": 0.25,
  "iou": 0.45
  }

• For small PPE items (like gloves), testing with imgsz=1280 in both Preview and API often lifts confidences. Just make sure both sides use the same value.

Where to pick the checkpoint in HUB

• Go to your model page → Deploy tab.
• In the Endpoint card, use the “Selected checkpoint” dropdown to choose Best (or a specific epoch).
• Click Stop Endpoint, then Start Endpoint to ensure the Dedicated endpoint reloads that exact artifact.
• If you don’t see the checkpoint selector, please share a screenshot of the Deploy tab so we can guide you to the exact control in your workspace.

If a gap persists after matching imgsz/conf/iou and checkpoint:

• Please capture:
  • A screenshot of Deploy → Selected checkpoint.
  • The Preview settings (imgsz/conf/iou) and a screenshot of the same detection with its confidence.
  • The API request (redact key) and the API JSON for that same image.
• With that, I’ll reproduce your exact path and pinpoint the cause quickly.

References

• See the Ultralytics HUB Inference API guide for request examples and argument details.
• For context on why deployment mismatches can happen (resize, NMS, byte differences), see Best Practices for Model Deployment.

You’re on the right track—once imgsz/conf/iou and the checkpoint are aligned, Preview and Dedicated API should match closely. Happy to keep digging with the screenshots/details above ✅

[vipin-prabhakaran]

hi @pderrenger we tried the all the options u told ,but still getting noticable difference in confidence values (as per requirement we can neglect till 2%). the bellow is the script i tried import json
import io,base64
import requests

Run inference on an image

url = "https://predict.ultralytics.com"
headers = {"x-api-key": "myapikey"}
data = {"model": "https://hub.ultralytics.com/models/modelid", "imgsz": 640, "conf": 0.25, "iou": 0.45}
with open("/path/grp5.png", "rb") as f:
response = requests.post(url, headers=headers, data=data, files={"file": f})

Check for successful response

response.raise_for_status()

Print inference results

print(json.dumps(response.json(), indent=2)). and this is giving a result like {
"images": [
{
"results": [
{
"box": {
"x1": 745.75464,
"x2": 1085.06458,
"y1": 388.19571,
"y2": 772.09497
},
"class": 8,
"confidence": 0.93324,
"name": "no_jacket"
},
{
"box": {
"x1": 503.69687,
"x2": 779.08575,
"y1": 374.93274,
"y2": 753.14038
},
"class": 8,
"confidence": 0.91877,
"name": "no_jacket"
},
{
"box": {
"x1": 180.85065,
"x2": 491.36661,
"y1": 404.13504,
"y2": 774.75977
},
"class": 8,
"confidence": 0.91448,
"name": "no_jacket"
},
{
"box": {
"x1": 1109.42188,
"x2": 1407.45691,
"y1": 402.93466,
"y2": 785.68488
},
"class": 10,
"confidence": 0.91129,
"name": "rf_jacket"
},
{
"box": {
"x1": 169.73305,
"x2": 328.76434,
"y1": 668.11371,
"y2": 934.28802
},
"class": 0,
"confidence": 0.8824,
"name": "gloves"
},
{
"box": {
"x1": 277.01968,
"x2": 406.69186,
"y1": 228.91498,
"y2": 398.90329
},
"class": 7,
"confidence": 0.86534,
"name": "no_helmet"
},
{
"box": {
"x1": 735.08124,
"x2": 829.44275,
"y1": 637.46613,
"y2": 868.61548
},
"class": 0,
"confidence": 0.84209,
"name": "gloves"
},
{
"box": {
"x1": 993.48364,
"x2": 1098.86511,
"y1": 622.1048,
"y2": 860.97235
},
"class": 0,
"confidence": 0.83125,
"name": "gloves"
},
{
"box": {
"x1": 679.94,
"x2": 769.65942,
"y1": 1111.88611,
"y2": 1216.42163
},
"class": 11,
"confidence": 0.78901,
"name": "shoes"
},
{
"box": {
"x1": 404.42786,
"x2": 544.74854,
"y1": 1131.59607,
"y2": 1254.29102
},
"class": 11,
"confidence": 0.7804,
"name": "shoes"
},
{
"box": {
"x1": 1272.21375,
"x2": 1366.08289,
"y1": 1123.47888,
"y2": 1210.34326
},
"class": 9,
"confidence": 0.77892,
"name": "no_shoes"
},
{
"box": {
"x1": 443.49579,
"x2": 529.16309,
"y1": 627.39417,
"y2": 881.07391
},
"class": 0,
"confidence": 0.76684,
"name": "gloves"
},
{
"box": {
"x1": 792.61792,
"x2": 887.31885,
"y1": 1098.54749,
"y2": 1224.76062
},
"class": 9,
"confidence": 0.74522,
"name": "no_shoes"
},
{
"box": {
"x1": 906.77094,
"x2": 1003.40399,
"y1": 1096.2804,
"y2": 1209.58203
},
"class": 9,
"confidence": 0.72996,
"name": "no_shoes"
},
{
"box": {
"x1": 1190.68323,
"x2": 1309.15271,
"y1": 259.0235,
"y2": 412.96106
},
"class": 5,
"confidence": 0.64346,
"name": "helmet_yellow"
},
{
"box": {
"x1": 579.8902,
"x2": 686.60022,
"y1": 1117.85559,
"y2": 1226.198
},
"class": 11,
"confidence": 0.58892,
"name": "shoes"
},
{
"box": {
"x1": 601.47168,
"x2": 722.98364,
"y1": 205.31317,
"y2": 388.49677
},
"class": 5,
"confidence": 0.58354,
"name": "helmet_yellow"
},
{
"box": {
"x1": 324.46552,
"x2": 413.09265,
"y1": 1148.55737,
"y2": 1293.29407
},
"class": 9,
"confidence": 0.55751,
"name": "no_shoes"
},
{
"box": {
"x1": 824.91608,
"x2": 950.27124,
"y1": 233.1595,
"y2": 403.05005
},
"class": 2,
"confidence": 0.54005,
"name": "helmet_blue"
},
{
"box": {
"x1": 323.3576,
"x2": 415.25339,
"y1": 1148.38892,
"y2": 1292.79736
},
"class": 11,
"confidence": 0.5198,
"name": "shoes"
},
{
"box": {
"x1": 1331.83154,
"x2": 1412.34485,
"y1": 582.73828,
"y2": 830.513
},
"class": 6,
"confidence": 0.4465,
"name": "no_gloves"
},
{
"box": {
"x1": 1121.34314,
"x2": 1187.20142,
"y1": 1125.7572,
"y2": 1215.36938
},
"class": 9,
"confidence": 0.38994,
"name": "no_shoes"
},
{
"box": {
"x1": 825.60022,
"x2": 949.40741,
"y1": 233.6897,
"y2": 403.10031
},
"class": 7,
"confidence": 0.35763,
"name": "no_helmet"
},
{
"box": {
"x1": 1118.69788,
"x2": 1189.24414,
"y1": 1124.3103,
"y2": 1216.59668
},
"class": 11,
"confidence": 0.30376,
"name": "shoes"
}
],
"shape": [
1328,
1770
],
"speed": {
"inference": 687.46101,
"postprocess": 0.74014,
"preprocess": 3.17581
}
}
],
"metadata": {
"functionTimeAlive": 42087.042,
"functionTimeCall": 1.851,
"imageCount": 1,
"model": "https://hub.ultralytics.com/models/modelid",
"version": {
"python": "3.11.10",
"torch": "2.8.0+cpu",
"torchvision": "0.23.0+cpu",
"ultralytics": "8.3.189"
}
}
}. u can check the helmet_yellow class to notice the difference quickly ,we cant tell the client that the detction differences in preview and our app as we are relying heavily on the api results .this is just a sample ,its happening across all images .bellow is the preview result of the same image

. and the the image also i am giving here for your refference as u asked

. the deploy screen shot of my hub page is

,we need to know the root cause of the difference as our client always compare it with preview section results in hub ,i hope you will understand our desperation to find the root cause ,as we have given all the details ,we really look forword for a solution .

[pderrenger]

Hi @vipin-prabhakaran — thanks for sharing the image, JSON, and screenshots. Super helpful. I know how important it is to keep API and Preview aligned for client reviews — we’ll get this pinned down. 🔍

What most likely explains the gap you’re seeing

• IoU threshold mismatch (most common): HUB Preview uses the Ultralytics Python engine defaults, where IoU for NMS is 0.70 by default, while the Inference API defaults to 0.45 unless you override it. That alone can easily swing which candidate box survives NMS and thus the displayed confidence. Try forcing iou=0.70 in your API call and match the Preview slider to 0.70. See default settings in the Ultralytics configuration reference. See the Configuration guide for default inference settings.
• Minor engine/version drift: Your API metadata shows ultralytics 8.3.189 (CPU). A different backend build in Preview can cause tiny (<1–2%) numeric differences, but not the larger swings you described; IoU mismatches are the usual culprit.

Quick test to confirm (using your grp5.png)

• In Preview: set imgsz=640, conf=0.25, IoU=0.70.
• In API: send exactly the same bytes and force iou=0.70.

Minimal API example

• Python

  import requests, io, json

  url = "https://predict.ultralytics.com" # or your Dedicated endpoint
  headers = {"x-api-key": "YOUR_KEY"}
  data = {
  "model": "https://hub.ultralytics.com/models/modelid", # your MODEL_ID
  "imgsz": 640,
  "conf": 0.25,
  "iou": 0.70, # match Preview
  }
  with open("/path/grp5.png", "rb") as f:
  r = requests.post(url, headers=headers, data=data, files={"file": ("grp5.png", io.BytesIO(f.read()))})
  r.raise_for_status()
  print(json.dumps(r.json(), indent=2))

How we can quickly isolate any remaining deltas

• Verify we’re comparing the same detection: Normalize the box you care about from both sides and compare. For example, x1n = x1 / width, y1n = y1 / height, etc., using the “shape” returned in the API JSON. That ensures we’re looking at the same box, not a neighbor after NMS. See the Results reference on normalized coordinates.
• Capture the Preview JSON once: In your browser’s DevTools (Network tab) while using Preview, locate the predict call, open the Response, and copy the JSON (redact keys). This will show the exact Preview params and results so we can diff against your API JSON line-by-line.
• Ensure the exact checkpoint is used: On your model page → Deploy tab → Selected checkpoint = Best (or the specific epoch you want), then Stop Endpoint → Start Endpoint to reload for Dedicated endpoints. Preview uses the same selected checkpoint. If you don’t see the selector, share a screenshot and we’ll point it out.

Why IoU matters here

• IoU controls suppression in NMS. At 0.45, more boxes are suppressed earlier; at 0.70, fewer are suppressed, which can change which candidate box “wins” and its associated confidence. This looks like a confidence change, but it’s really a different survivor from the candidate set. See the Ultralytics HUB Inference API guide for API arguments and defaults and the Configuration guide for default inference settings.

If you can:

• Re-run your grp5.png with iou=0.70 on both Preview and API.
• Post the Preview JSON (copied from DevTools) and the API JSON for that same run.
• Call out the exact detection you’re using to compare (e.g., helmet_yellow at [x1,y1,x2,y2]).

With those two JSONs side-by-side, we can confirm alignment (my expectation is you’ll fall within your 2% tolerance once IoU is matched). If not, I’ll escalate with engineering using your artifacts so we can resolve any remaining backend drift.

References

• Use the Ultralytics HUB Inference API and available arguments, including imgsz/conf/iou: see the Ultralytics HUB Inference API guide.
• Default inference settings (e.g., IoU default 0.70 in the Python engine): see the Configuration guide.
• Tips to align deployment behavior across environments (resize/letterbox, NMS, engine differences): see Best Practices for Model Deployment.

Appreciate your patience and the thorough data — once we lock IoU and checkpoint, Preview and API should match very closely. ✅

[vipin-prabhakaran]

hi @pderrenger thanks for your response ,i tried with iou .70 on both api and preview ,but still i am getting the response with a difference as preview showing helmet_yellow as 25% and its correct, but the api giving helmet_yellow as both 58% and 64% which will make a case for false detection,this much difference we cannot explain to a client ,i hope there should be a proper root cause fix to this ,and it will be highly helpfull for our application
Any help to find and fix the root cause of this will be greatly appreciated ,thanks in advance

[pderrenger]

Hi @vipin-prabhakaran — thanks for sticking with this and for the thorough artifacts. I agree: a swing like ~25% (Preview) vs ~58–64% (API) on the same “helmet_yellow” is beyond acceptable drift, so let’s pin this down.

What we need to verify next (likely cause + how to prove it)

• Verify we are comparing the same candidate box. With IoU=0.70, NMS leaves more overlapping candidates. It’s common for the UI to highlight a different survivor (with a different score) than the one you’re reading in the API JSON. The only way to be 100% sure is to compare normalized box coordinates from both sides.
• Action: Please capture the Preview’s raw JSON once so we can diff box-for-box.
  • Open the model Preview, set imgsz=640, conf=0.25, iou=0.70.
  • Run the same image.
  • In your browser DevTools → Network, open the predict call and copy the full Response JSON (redact keys). This also shows the exact params Preview used.
  • Post that JSON plus the API JSON you already shared for the same image.

How to confirm same box vs a neighbor (quick check)

• From each JSON, normalize the box coords using the “shape” field. If the normalized boxes match within ~1e-3, it’s the same detection; if they differ, we’ve been comparing different survivors of NMS.

Minimal helper to normalize (run locally on your API JSON)

• Python

  import json

  def norm(box, shape):
  h, w = shape
  return {
  "x1n": box["x1"] / w,
  "y1n": box["y1"] / h,
  "x2n": box["x2"] / w,
  "y2n": box["y2"] / h,
  }

  paste your API JSON into resp

  resp = json.loads(...)

  img = resp["images"][0]
  shape = img["shape"] # [h, w]
  for det in img["results"]:
  if det["name"] == "helmet_yellow":
  print(det["confidence"], norm(det["box"], shape))

Please run the same on the Preview JSON once you share it; I’ll do a line-by-line match and point to the exact detection pair.

Hard-pin everything we can control (one more time)

• Params: Keep imgsz/conf/iou identical in both places. On API, include iou=0.70 explicitly to match Preview. See available arguments in the Ultralytics HUB Inference API guide. See the Ultralytics Docs for details.
• Defaults nuance: The Python engine default IoU is 0.70 (Preview), while the API default IoU is 0.45 unless you override it. You’ve already set 0.70 on both, which is correct. See Configuration - Predict Settings.
• Checkpoint: Please confirm Deploy → Selected checkpoint is Best (or the exact epoch you intend), then Stop Endpoint → Start Endpoint to reload. Preview uses the selected checkpoint too. If the dropdown isn’t visible in your workspace, share a fresh Deploy screenshot and I’ll point out where it should be.
• Byte identity: The API “shape” you posted is [1328, 1770]. Please confirm Preview shows that same input size for grp5.png (hover or check DevTools request payload). Any re-save/EXIF change will alter confidences.

Two quick isolation tests (optional but very telling)

• Crop test (removes NMS interactions): Crop a tight patch around the helmet and run both Preview and API at iou=0.70. If scores now match, the earlier gap was due to different NMS survivors in a crowded region.
• Shared vs Dedicated parity: You already tested Shared (8.3.189 CPU). If your Dedicated endpoint is still enabled, run the same bytes through it with identical params and share both JSONs. This will rule out any backend variance.

Why we’re asking for the Preview JSON

• It’s the only artifact that lets us confirm whether the UI’s highlighted “helmet_yellow 25%” is the same candidate as the API’s 58–64% boxes, or a different overlapping box left unsuppressed at IoU=0.70. Once we have both JSONs, we can either (a) show the two boxes differ and tune NMS accordingly, or (b) escalate an engine/UI discrepancy with a minimal repro.

References

• Inference parameters and examples: see the Ultralytics HUB Inference API guide.
• Default inference settings (e.g., IoU=0.70 for the Python engine): see Configuration – Predict Settings.
• Why small preprocessing/NMS changes cause output shifts and how to align environments: see Best Practices for Model Deployment.

If you can share the Preview JSON for grp5.png at imgsz=640, conf=0.25, iou=0.70, I’ll diff it against your API JSON immediately and come back with a concrete root-cause and fix path. We’ll get this within your 2% tolerance. 🙏

[vipin-prabhakaran]

@pderrenger i am sharing the Preview JSON for grp5.png at imgsz=640, conf=0.25, iou=0.70, against my API JSON .i can clearly see the difference in shapes and difference in size in preview and my api although i have passed same grp5.png to both ,this is exactly what i asked in my first question itself do we need to resize before passing to api ,but as per you api is doing it internaly ,the bellow one is my json response from api {
"images": [
{
"results": [
{
"box": {
"x1": 745.75464,
"x2": 1085.06458,
"y1": 388.19571,
"y2": 772.09497
},
"class": 8,
"confidence": 0.93324,
"name": "no_jacket"
},
{
"box": {
"x1": 503.69687,
"x2": 779.08575,
"y1": 374.93274,
"y2": 753.14038
},
"class": 8,
"confidence": 0.91877,
"name": "no_jacket"
},
{
"box": {
"x1": 180.85065,
"x2": 491.36661,
"y1": 404.13504,
"y2": 774.75977
},
"class": 8,
"confidence": 0.91448,
"name": "no_jacket"
},
{
"box": {
"x1": 1109.42188,
"x2": 1407.45691,
"y1": 402.93466,
"y2": 785.68488
},
"class": 10,
"confidence": 0.91129,
"name": "rf_jacket"
},
{
"box": {
"x1": 169.73305,
"x2": 328.76434,
"y1": 668.11371,
"y2": 934.28802
},
"class": 0,
"confidence": 0.8824,
"name": "gloves"
},
{
"box": {
"x1": 277.01968,
"x2": 406.69186,
"y1": 228.91498,
"y2": 398.90329
},
"class": 7,
"confidence": 0.86534,
"name": "no_helmet"
},
{
"box": {
"x1": 735.08124,
"x2": 829.44275,
"y1": 637.46613,
"y2": 868.61548
},
"class": 0,
"confidence": 0.84209,
"name": "gloves"
},
{
"box": {
"x1": 993.48364,
"x2": 1098.86511,
"y1": 622.1048,
"y2": 860.97235
},
"class": 0,
"confidence": 0.83125,
"name": "gloves"
},
{
"box": {
"x1": 679.94,
"x2": 769.65942,
"y1": 1111.88611,
"y2": 1216.42163
},
"class": 11,
"confidence": 0.78901,
"name": "shoes"
},
{
"box": {
"x1": 404.42786,
"x2": 544.74854,
"y1": 1131.59607,
"y2": 1254.29102
},
"class": 11,
"confidence": 0.7804,
"name": "shoes"
},
{
"box": {
"x1": 1272.21375,
"x2": 1366.08289,
"y1": 1123.47888,
"y2": 1210.34326
},
"class": 9,
"confidence": 0.77892,
"name": "no_shoes"
},
{
"box": {
"x1": 443.49579,
"x2": 529.16309,
"y1": 627.39417,
"y2": 881.07391
},
"class": 0,
"confidence": 0.76684,
"name": "gloves"
},
{
"box": {
"x1": 792.61792,
"x2": 887.31885,
"y1": 1098.54749,
"y2": 1224.76062
},
"class": 9,
"confidence": 0.74522,
"name": "no_shoes"
},
{
"box": {
"x1": 906.77094,
"x2": 1003.40399,
"y1": 1096.2804,
"y2": 1209.58203
},
"class": 9,
"confidence": 0.72996,
"name": "no_shoes"
},
{
"box": {
"x1": 1190.68323,
"x2": 1309.15271,
"y1": 259.0235,
"y2": 412.96106
},
"class": 5,
"confidence": 0.64346,
"name": "helmet_yellow"
},
{
"box": {
"x1": 579.8902,
"x2": 686.60022,
"y1": 1117.85559,
"y2": 1226.198
},
"class": 11,
"confidence": 0.58892,
"name": "shoes"
},
{
"box": {
"x1": 601.47168,
"x2": 722.98364,
"y1": 205.31317,
"y2": 388.49677
},
"class": 5,
"confidence": 0.58354,
"name": "helmet_yellow"
},
{
"box": {
"x1": 324.46552,
"x2": 413.09265,
"y1": 1148.55737,
"y2": 1293.29407
},
"class": 9,
"confidence": 0.55751,
"name": "no_shoes"
},
{
"box": {
"x1": 824.91608,
"x2": 950.27124,
"y1": 233.1595,
"y2": 403.05005
},
"class": 2,
"confidence": 0.54005,
"name": "helmet_blue"
},
{
"box": {
"x1": 323.3576,
"x2": 415.25339,
"y1": 1148.38892,
"y2": 1292.79736
},
"class": 11,
"confidence": 0.5198,
"name": "shoes"
},
{
"box": {
"x1": 1331.83154,
"x2": 1412.34485,
"y1": 582.73828,
"y2": 830.513
},
"class": 6,
"confidence": 0.4465,
"name": "no_gloves"
},
{
"box": {
"x1": 1121.34314,
"x2": 1187.20142,
"y1": 1125.7572,
"y2": 1215.36938
},
"class": 9,
"confidence": 0.38994,
"name": "no_shoes"
},
{
"box": {
"x1": 825.60022,
"x2": 949.40741,
"y1": 233.6897,
"y2": 403.10031
},
"class": 7,
"confidence": 0.35763,
"name": "no_helmet"
},
{
"box": {
"x1": 1118.69788,
"x2": 1189.24414,
"y1": 1124.3103,
"y2": 1216.59668
},
"class": 11,
"confidence": 0.30376,
"name": "shoes"
}
],
"shape": [
1328,
1770
],
"speed": {
"inference": 1624.64841,
"postprocess": 58.49561,
"preprocess": 3.05553
}
}
],
"metadata": {
"functionTimeAlive": 538.217,
"functionTimeCall": 1.779,
"imageCount": 1,
"model": "https://hub.ultralytics.com/models/myid”,
"version": {
"python": "3.11.10",
"torch": "2.8.0+cpu",
"torchvision": "0.23.0+cpu",
"ultralytics": "8.3.183"
}
}. and json from preview clearly showing a resized image and change in shapes in class box is evident .the json from preview with iou .70 is giving like this {
"images": [
{
"results": [
{
"box": {
"x1": 269.41626,
"x2": 391.92407,
"y1": 140.93262,
"y2": 278.8833
},
"class": 8,
"confidence": 0.93239,
"name": "no_jacket"
},
{
"box": {
"x1": 181.75809,
"x2": 281.50476,
"y1": 135.76315,
"y2": 272.4292
},
"class": 8,
"confidence": 0.91849,
"name": "no_jacket"
},
{
"box": {
"x1": 64.97681,
"x2": 177.27448,
"y1": 146.30737,
"y2": 280.32132
},
"class": 8,
"confidence": 0.91134,
"name": "no_jacket"
},
{
"box": {
"x1": 401.4801,
"x2": 508.04395,
"y1": 145.78674,
"y2": 284.56757
},
"class": 10,
"confidence": 0.9054,
"name": "rf_jacket"
},
{
"box": {
"x1": 60.67363,
"x2": 119.18661,
"y1": 241.06815,
"y2": 338.71768
},
"class": 0,
"confidence": 0.88018,
"name": "gloves"
},
{
"box": {
"x1": 265.28009,
"x2": 300.04871,
"y1": 230.18364,
"y2": 314.52997
},
"class": 0,
"confidence": 0.84721,
"name": "gloves"
},
{
"box": {
"x1": 99.58792,
"x2": 148.18082,
"y1": 82.00653,
"y2": 145.54898
},
"class": 7,
"confidence": 0.83543,
"name": "no_helmet"
},
{
"box": {
"x1": 358.88043,
"x2": 397.37958,
"y1": 224.42375,
"y2": 311.8476
},
"class": 0,
"confidence": 0.8324,
"name": "gloves"
},
{
"box": {
"x1": 146.61604,
"x2": 197.4511,
"y1": 408.55423,
"y2": 454.116
},
"class": 11,
"confidence": 0.79681,
"name": "shoes"
},
{
"box": {
"x1": 160.07397,
"x2": 191.28491,
"y1": 225.99899,
"y2": 318.53549
},
"class": 0,
"confidence": 0.77428,
"name": "gloves"
},
{
"box": {
"x1": 245.58965,
"x2": 278.65302,
"y1": 401.64252,
"y2": 440.83331
},
"class": 11,
"confidence": 0.7673,
"name": "shoes"
},
{
"box": {
"x1": 459.64197,
"x2": 494.33289,
"y1": 405.39267,
"y2": 438.16653
},
"class": 9,
"confidence": 0.7543,
"name": "no_shoes"
},
{
"box": {
"x1": 286.31323,
"x2": 321.73669,
"y1": 396.3476,
"y2": 443.63184
},
"class": 9,
"confidence": 0.74613,
"name": "no_shoes"
},
{
"box": {
"x1": 116.64233,
"x2": 150.61005,
"y1": 414.97919,
"y2": 467.94159
},
"class": 11,
"confidence": 0.72687,
"name": "shoes"
},
{
"box": {
"x1": 297.99786,
"x2": 343.89514,
"y1": 83.94962,
"y2": 146.18411
},
"class": 7,
"confidence": 0.6708,
"name": "no_helmet"
},
{
"box": {
"x1": 326.97406,
"x2": 364.03986,
"y1": 395.8042,
"y2": 438.46918
},
"class": 9,
"confidence": 0.66638,
"name": "no_shoes"
},
{
"box": {
"x1": 429.59793,
"x2": 473.72354,
"y1": 93.3277,
"y2": 150.17686
},
"class": 7,
"confidence": 0.64814,
"name": "no_helmet"
},
{
"box": {
"x1": 209.02774,
"x2": 248.54648,
"y1": 403.45157,
"y2": 444.49014
},
"class": 11,
"confidence": 0.46042,
"name": "shoes"
},
{
"box": {
"x1": 481.04453,
"x2": 511.0303,
"y1": 210.82477,
"y2": 301.11215
},
"class": 6,
"confidence": 0.26189,
"name": "no_gloves"
},
{
"box": {
"x1": 216.87207,
"x2": 261.43176,
"y1": 73.72946,
"y2": 140.18939
},
"class": 5,
"confidence": 0.25794,
"name": "helmet_yellow"
}
],
"shape": [
480,
640
],
"speed": {
"inference": 624.65736,
"postprocess": 6.03301,
"preprocess": 2.33054
}
}
],
"metadata": {
"functionTimeAlive": 18195.386,
"functionTimeCall": 0.642,
"imageCount": 1,
"model": "https://hub.ultralytics.com/models/id”,
"version": {
"python": "3.11.10",
"torch": "2.8.0+cpu",
"torchvision": "0.23.0+cpu",
"ultralytics": "8.3.191"
}
}
}. now i hope u will give me a proper solution than suggesting some other trial and error methode as the root cause could have been clear to you .i am seing shape in preview is not having full image shape as it is in api ,from preview json i can clearly see image shape is 640 * 480 as in my api its 1328 * 1770 .clearly this resize algorithem is not working in api case .i hope your team will close this time is getting delayed for me to get the confidence back from my client

[pderrenger]

Hi @vipin-prabhakaran — thanks for sharing both JSON payloads. You’ve surfaced the root cause clearly:

What’s happening

• The Preview request is sending a browser-resized copy of your image bounded by imgsz, so the model actually sees 480×640 there. That’s why the Preview JSON reports "shape": [480, 640].
• Your API call is sending the original 1328×1770 bytes, so the API JSON shows "shape": [1328, 1770]. The model is therefore seeing different pixels across the two paths, which can easily shift confidences on small PPE items (e.g., helmet colors). The “shape” field reflects the original image size used by the engine, not the internal network tensor, so seeing different shapes confirms different input bytes. See Results’ orig_shape/shape behavior in the docs. (docs.ultralytics.com)

How to make API match Preview exactly (quick fix you can ship)

• Downscale on your side to match Preview’s upload behavior: constrain the longer side to imgsz (e.g., 640), keep aspect ratio, then post those bytes to the API. After this, the API JSON shape should read [480, 640] for your grp5.png and confidences should align with Preview within your tolerance.

Minimal Python example (keeps PNG to avoid extra JPEG artifacts)

• Make sure you also set the same iou as Preview (Ultralytics Python engine default is 0.70 unless you override it). (docs.ultralytics.com)

  import io
  from PIL import Image
  import requests

  path = "/path/grp5.png"
  url = "https://predict.ultralytics.com" # or your Dedicated endpoint
  headers = {"x-api-key": "YOUR_KEY"}
  data = {
  "model": "https://hub.ultralytics.com/models/YOUR_MODEL_ID",
  "imgsz": 640,
  "conf": 0.25,
  "iou": 0.70, # match Preview
  }

  Resize so max(h, w) == imgsz, keep aspect (Preview behavior)

  imgsz = 640
  im = Image.open(path).convert("RGB")
  w, h = im.size
  scale = imgsz / max(w, h)
  nw, nh = int(round(w * scale)), int(round(h * scale))
  im = im.resize((nw, nh), resample=Image.BILINEAR)

  buf = io.BytesIO()
  im.save(buf, format="PNG")
  buf.seek(0)

  r = requests.post(url, headers=headers, data=data, files={"file": ("grp5.png", buf)})
  print(r.json()["images"][0]["shape"]) # expect [480, 640]

Notes and next steps

• Arguments to keep identical: imgsz/conf/iou. The API accepts these as shown in the docs; using iou=0.70 in both places prevents NMS-driven drift. (docs.ultralytics.com)
• If you prefer to run end-to-end at higher detail, set imgsz=1280 in both Preview and API, and pre-resize to max side 1280 before posting to API (same code, just change imgsz).
• Minor engine version differences (you showed 8.3.183/189 vs 8.3.191) can create tiny numerical drift, but the large gap you observed is explained by different input bytes. We’ll keep this issue open and I’ve flagged an internal task to:
  1. unify Preview/API preprocessing, and
  2. expose an option in Preview to upload original bytes for exact parity.
• If you want me to double-check parity on your image after the pre-resize change, post the new API JSON and a fresh Preview JSON at imgsz=640, iou=0.70. I’ll verify box-for-box.

References

• Inference API request parameters and usage: See the Ultralytics HUB Inference API guide. (docs.ultralytics.com)
• Default inference settings (e.g., iou=0.70 in the Python engine): See Predict mode defaults. (docs.ultralytics.com)
• What “shape” represents in results and why it can differ: See Results reference (orig_shape/shape). (docs.ultralytics.com)

Appreciate your patience here — aligning the input bytes should get you within your 2% target quickly. 🙏

[vipin-prabhakaran]

@pderrenger this issue is not been solved even after i used your resize algorithem ,u can see the difference in json bellow
preview json
{
"images": [
{
"results": [
{
"box": {
"x1": 74.18079,
"x2": 421.84094,
"y1": 353.25714,
"y2": 799.13312
},
"class": 8,
"confidence": 0.97873,
"name": "no_jacket"
},
{
"box": {
"x1": 68.70674,
"x2": 133.4579,
"y1": 614.36212,
"y2": 851.89044
},
"class": 6,
"confidence": 0.94638,
"name": "no_gloves"
},
{
"box": {
"x1": 301.98584,
"x2": 410.12,
"y1": 1078.93457,
"y2": 1197.92383
},
"class": 11,
"confidence": 0.94322,
"name": "shoes"
},
{
"box": {
"x1": 580.31787,
"x2": 731.70142,
"y1": 68.43619,
"y2": 274.68643
},
"class": 4,
"confidence": 0.92422,
"name": "helmet_white"
},
{
"box": {
"x1": 358.25497,
"x2": 431.97672,
"y1": 600.56494,
"y2": 832.77832
},
"class": 0,
"confidence": 0.92101,
"name": "gloves"
},
{
"box": {
"x1": 684.48834,
"x2": 785.80267,
"y1": 1061.57861,
"y2": 1190.97412
},
"class": 11,
"confidence": 0.9071,
"name": "shoes"
},
{
"box": {
"x1": 165.55334,
"x2": 286.46719,
"y1": 1100.34058,
"y2": 1221.9104
},
"class": 11,
"confidence": 0.88142,
"name": "shoes"
},
{
"box": {
"x1": 504.42001,
"x2": 612.66052,
"y1": 1068.97754,
"y2": 1174.56641
},
"class": 11,
"confidence": 0.88012,
"name": "shoes"
},
{
"box": {
"x1": 68.61359,
"x2": 133.48184,
"y1": 614.2793,
"y2": 851.75171
},
"class": 0,
"confidence": 0.67913,
"name": "gloves"
},
{
"box": {
"x1": 782.92578,
"x2": 872.43555,
"y1": 578.15173,
"y2": 800.28503
},
"class": 6,
"confidence": 0.50932,
"name": "no_gloves"
},
{
"box": {
"x1": 493.40967,
"x2": 872.83545,
"y1": 256.95319,
"y2": 724.19971
},
"class": 8,
"confidence": 0.31538,
"name": "no_jacket"
}
],
"shape": [
1280,
960
],
"speed": {
"inference": 7277.18606,
"postprocess": 1.93673,
"preprocess": 5.57611
}
}
],
"metadata": {
"functionTimeAlive": 86283.319,
"functionTimeCall": 9.579,
"imageCount": 1,
"model": "https://hub.ultralytics.com/models/my id",
"version": {
"python": "3.11.10",
"torch": "2.8.0+cpu",
"torchvision": "0.23.0+cpu",
"ultralytics": "8.3.192"
}
}
}. and the api i used the resize script also as u suggested ,i am getting this json

{
"images": [
{
"results": [
{
"box": {
"x1": 73.93427,
"x2": 420.49707,
"y1": 352.99939,
"y2": 799.25415
},
"class": 8,
"confidence": 0.97929,
"name": "no_jacket"
},
{
"box": {
"x1": 302.41547,
"x2": 409.9693,
"y1": 1078.8988,
"y2": 1198.18079
},
"class": 11,
"confidence": 0.95342,
"name": "shoes"
},
{
"box": {
"x1": 68.46637,
"x2": 133.72968,
"y1": 613.53979,
"y2": 849.60181
},
"class": 6,
"confidence": 0.94377,
"name": "no_gloves"
},
{
"box": {
"x1": 358.30231,
"x2": 431.54858,
"y1": 600.01025,
"y2": 832.74805
},
"class": 0,
"confidence": 0.93309,
"name": "gloves"
},
{
"box": {
"x1": 580.20105,
"x2": 733.41174,
"y1": 68.14682,
"y2": 275.19727
},
"class": 5,
"confidence": 0.90669,
"name": "helmet_yellow"
},
{
"box": {
"x1": 684.64795,
"x2": 786.10938,
"y1": 1063.28979,
"y2": 1191.03833
},
"class": 11,
"confidence": 0.88352,
"name": "shoes"
},
{
"box": {
"x1": 165.47705,
"x2": 286.09418,
"y1": 1099.29602,
"y2": 1222.02722
},
"class": 11,
"confidence": 0.86566,
"name": "shoes"
},
{
"box": {
"x1": 785.43768,
"x2": 871.38348,
"y1": 575.31323,
"y2": 796.50891
},
"class": 6,
"confidence": 0.79975,
"name": "no_gloves"
},
{
"box": {
"x1": 504.45056,
"x2": 610.44299,
"y1": 1067.72681,
"y2": 1174.61108
},
"class": 11,
"confidence": 0.75303,
"name": "shoes"
},
{
"box": {
"x1": 149.24905,
"x2": 275.63156,
"y1": 183.87189,
"y2": 357.04333
},
"class": 7,
"confidence": 0.49611,
"name": "no_helmet"
}
],
"shape": [
1280,
960
],
"speed": {
"inference": 26122.04793,
"postprocess": 173.60695,
"preprocess": 5.08449
}
}
],
"metadata": {
"functionTimeAlive": 40.985,
"functionTimeCall": 26.411,
"imageCount": 1,
"model": "https://hub.ultralytics.com/models/my id ",
"version": {
"python": "3.11.10",
"torch": "2.8.0+cpu",
"torchvision": "0.23.0+cpu",
"ultralytics": "8.3.191"
}
}. i am not getting after doing all this along why we are not been to identify the root cause of this