Ultralytics YOLO27: Join the waitlist.
Get Started

YOLO-World Model#

The YOLO-World Model is a real-time, Ultralytics YOLOv8-based approach to Open-Vocabulary Detection. It detects objects in an image from descriptive text. By lowering computational demands while preserving competitive accuracy, YOLO-World suits a wide range of vision applications.



Watch: YOLO World training workflow on custom dataset

YOLO-World Model architecture overview

Overview#

YOLO-World addresses the challenges of traditional Open-Vocabulary detection models, which often rely on heavy Transformer models requiring extensive computational resources. Closed-set detectors, in turn, depend on predefined object categories, which restricts their use in dynamic scenarios. YOLO-World extends the YOLOv8 framework with open-vocabulary detection, using vision-language modeling and pre-training on large datasets to identify a broad range of objects in zero-shot scenarios efficiently.

For open-vocabulary work that also needs instance masks, visual prompts, or a prompt-free mode, see YOLOE, which keeps the same set_classes() API.

Key Features#

  1. Real-time Solution: Built on fast CNNs, YOLO-World delivers open-vocabulary detection quickly enough for applications that need immediate results.

  2. Efficiency and Performance: YOLO-World reduces computational and resource requirements while maintaining accuracy, offering an efficient alternative to heavier transformer-based open-vocabulary detectors and enabling real-time applications.

  3. Inference with Offline Vocabulary: YOLO-World introduces a "prompt-then-detect" strategy, using an offline vocabulary to further improve efficiency. This approach enables the use of custom prompts computed a priori, including captions or categories, to be encoded and stored as offline vocabulary embeddings, streamlining the detection process.

  4. Powered by YOLOv8: Built upon Ultralytics YOLOv8, YOLO-World adapts its real-time detection architecture for fast, accurate open-vocabulary detection.

  5. Benchmark Excellence: YOLO-World outperforms existing open-vocabulary detectors, including MDETR and the GLIP series, in speed and efficiency on standard benchmarks, measured on a single NVIDIA V100 GPU.

  6. Versatile Applications: YOLO-World's efficiency makes open-vocabulary detection practical for many vision tasks, running much faster than earlier open-vocabulary methods.

Available Models, Supported Tasks, and Operating Modes#

This section details the models available with their specific pretrained weights, the tasks they support, and their compatibility with various operating modes such as Training, Validation, Inference, and Export, denoted by ✅ for supported modes and ❌ for unsupported modes.

Note

All YOLOv8-World weights are migrated directly from the official YOLO-World repository; credit for them goes to its authors.

Model TypePretrained WeightsTasks SupportedTrainingValidationInferenceExport
YOLOv8s-worldyolov8s-world.ptObject Detection✅✅✅❌
YOLOv8s-worldv2yolov8s-worldv2.ptObject Detection✅✅✅✅
YOLOv8m-worldyolov8m-world.ptObject Detection✅✅✅❌
YOLOv8m-worldv2yolov8m-worldv2.ptObject Detection✅✅✅✅
YOLOv8l-worldyolov8l-world.ptObject Detection✅✅✅❌
YOLOv8l-worldv2yolov8l-worldv2.ptObject Detection✅✅✅✅
YOLOv8x-worldyolov8x-world.ptObject Detection✅✅✅❌
YOLOv8x-worldv2yolov8x-worldv2.ptObject Detection✅✅✅✅

Zero-shot Transfer on COCO Dataset#

Performance
Model TypemAP50-95mAP50mAP75
yolov8s-world37.452.040.6
yolov8s-worldv237.752.241.0
yolov8m-world42.057.045.6
yolov8m-worldv243.058.446.8
yolov8l-world45.761.349.8
yolov8l-worldv245.861.349.8
yolov8x-world47.063.051.2
yolov8x-worldv247.162.851.4

Usage Examples#

Ultralytics provides a Python API and CLI commands for YOLO-World models.



Watch: YOLO-World Model Usage examples with Ultralytics | Open Vocab, Prompt-Free & others

Train Usage#

Tip

We strongly recommend using yolov8-worldv2 for custom training because it supports deterministic training and exports more easily to formats such as ONNX and TensorRT.

Train a YOLO-World object detection model with the train method, shown below on the COCO8 example dataset:

Example

PyTorch pretrained *.pt models as well as configuration *.yaml files can be passed to the YOLOWorld() class to create a model instance in Python:

from ultralytics import YOLOWorld

# Load a pretrained YOLOv8s-worldv2 model
model = YOLOWorld("yolov8s-worldv2.pt")

# Train the model on the COCO8 example dataset for 100 epochs
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

# Run inference with the YOLO-World model on the 'bus.jpg' image
results = model("path/to/bus.jpg")

Predict Usage#

Run object detection with the predict method, as shown below:

Example
from ultralytics import YOLOWorld

# Initialize a YOLO-World model
model = YOLOWorld("yolov8s-world.pt")  # or select yolov8m/l-world.pt for different sizes

# Execute inference with the YOLOv8s-world model on the specified image
results = model.predict("path/to/image.jpg")

# Show results
results[0].show()

This snippet loads a pretrained model and runs prediction on an image.

Val Usage#

Validate the model on a dataset as follows:

Example
from ultralytics import YOLO

# Create a YOLO-World model
model = YOLO("yolov8s-world.pt")  # or select yolov8m/l-world.pt for different sizes

# Conduct model validation on the COCO8 example dataset
metrics = model.val(data="coco8.yaml")

Track Usage#

Run object tracking with a YOLO-World model on videos or images as follows:

Example
from ultralytics import YOLO

# Create a YOLO-World model
model = YOLO("yolov8s-world.pt")  # or select yolov8m/l-world.pt for different sizes

# Track with a YOLO-World model on a video
results = model.track(source="path/to/video.mp4")
Note

The YOLO-World models provided by Ultralytics come pre-configured with COCO dataset categories as part of their offline vocabulary, so they are ready to use immediately. This allows the YOLOv8-World models to directly recognize and predict the 80 standard categories defined in the COCO dataset without requiring additional setup or customization.

Set prompts#

YOLO-World prompt class names overview

The YOLO-World framework allows dynamic specification of classes through custom prompts, letting users tailor the model to their specific needs without retraining. This feature is particularly useful for adapting the model to new domains or specific tasks that were not originally part of the training data. By setting custom prompts, users guide the model toward objects of interest, improving the relevance and accuracy of the detection results.

For instance, if your application only requires detecting 'person' and 'bus' objects, you can specify these classes directly:

Example
from ultralytics import YOLO

# Initialize a YOLO-World model
model = YOLO("yolov8s-world.pt")  # or choose yolov8m/l-world.pt

# Define custom classes
model.set_classes(["person", "bus"])

# Execute prediction for specified categories on an image
results = model.predict("path/to/image.jpg")

# Show results
results[0].show()
Background Class

Some users have found that appending an empty string "" as a background class can improve detection performance in certain scenarios. This behavior appears to be scenario-dependent and the exact mechanism is not fully understood:

model.set_classes(["person", "bus", ""])

The empty string stays in model.names as a class of its own, including in any model you save, so its detections come back with an empty label. Pass classes=[0, 1] at prediction time to keep only your real classes.

You can also save a model after setting custom classes. By doing this you create a version of the YOLO-World model that is specialized for your specific use case. This process embeds your custom class definitions directly into the model file, making the model ready to use with your specified classes without further adjustments. Follow these steps to save and load your custom YOLO-World model:

Example

First load a YOLO-World model, set custom classes for it and save it:

from ultralytics import YOLO

# Initialize a YOLO-World model
model = YOLO("yolov8s-world.pt")  # or select yolov8m/l-world.pt

# Define custom classes
model.set_classes(["person", "bus"])

# Save the model with the defined offline vocabulary
model.save("custom_yolov8s.pt")

After saving, the custom_yolov8s.pt model behaves like any other pretrained YOLOv8 model but with a key difference: it is now optimized to detect only the classes you have defined. This customization can significantly improve detection performance and efficiency for your specific application scenarios.

from ultralytics import YOLO

# Load your custom model
model = YOLO("custom_yolov8s.pt")

# Run inference to detect your custom classes
results = model.predict("path/to/image.jpg")

# Show results
results[0].show()

Benefits of Saving with Custom Vocabulary#

  • Efficiency: Streamlines the detection process by focusing on relevant objects, reducing computational overhead and speeding up inference.
  • Flexibility: Allows for easy adaptation of the model to new or niche detection tasks without the need for extensive retraining or data collection.
  • Simplicity: Simplifies deployment by eliminating the need to repeatedly specify custom classes at runtime, making the model directly usable with its embedded vocabulary.
  • Performance: Enhances detection accuracy for specified classes by focusing the model's attention and resources on recognizing the defined objects.

This approach customizes object detection models for specific tasks without retraining, making open-vocabulary detection practical for a broader range of applications.

Reproduce official results from scratch (Experimental)#

Prepare datasets#

  • Train data
DatasetTypeSamplesBoxesAnnotation Files
Objects365v1Detection609k9621kobjects365_train.json
GQAGrounding621k3681kfinal_mixed_train_no_coco.json
Flickr30kGrounding149k641kfinal_flickr_separateGT_train.json
  • Val data
DatasetTypeAnnotation Files
LVIS minivalDetectionminival.txt

Launch training from scratch#

Note

WorldTrainerFromScratch is highly customized to allow training YOLO-World models on both detection datasets and grounding datasets simultaneously.

Example
from ultralytics import YOLOWorld
from ultralytics.models.yolo.world.train_world import WorldTrainerFromScratch

# Option 1: Use Python dictionary
data = {
    "train": {
        "yolo_data": ["Objects365.yaml"],
        "grounding_data": [
            {
                "img_path": "flickr30k/images",
                "json_file": "flickr30k/final_flickr_separateGT_train.json",
            },
            {
                "img_path": "GQA/images",
                "json_file": "GQA/final_mixed_train_no_coco.json",
            },
        ],
    },
    "val": {"yolo_data": ["lvis.yaml"]},
}

# Option 2: Use YAML file (yolo_world_data.yaml)
# train:
#   yolo_data:
#     - Objects365.yaml
#   grounding_data:
#     - img_path: flickr/full_images/
#       json_file: flickr/annotations/final_flickr_separateGT_train_segm.json
#     - img_path: mixed_grounding/gqa/images
#       json_file: mixed_grounding/annotations/final_mixed_train_no_coco_segm.json
# val:
#   yolo_data:
#     - lvis.yaml

model = YOLOWorld("yolov8s-worldv2.yaml")
model.train(
    data=data,  # or data="yolo_world_data.yaml" if using YAML file
    batch=128,
    epochs=100,
    trainer=WorldTrainerFromScratch,
)

Citations and Acknowledgments#

We extend our gratitude to the Tencent AILab Computer Vision Center for their pioneering work in real-time open-vocabulary object detection with YOLO-World:

Citation
@article{cheng2024yolow,
title={YOLO-World: Real-Time Open-Vocabulary Object Detection},
author={Cheng, Tianheng and Song, Lin and Ge, Yixiao and Liu, Wenyu and Wang, Xinggang and Shan, Ying},
journal={arXiv preprint arXiv:2401.17270},
year={2024}
}

For further reading, the original YOLO-World paper is available on arXiv. The project's source code and additional resources can be accessed via their GitHub repository. We appreciate their commitment to advancing the field and sharing their valuable insights with the community.

FAQ#

  • The YOLO-World model is a real-time object detection approach based on the Ultralytics YOLOv8 framework. It performs Open-Vocabulary Detection by identifying objects within an image based on descriptive text. Using vision-language modeling and pre-training on large datasets, YOLO-World achieves high efficiency with reduced computational demands, making it well suited for real-time applications.

  • YOLO-World supports a "prompt-then-detect" strategy, which uses an offline vocabulary to improve efficiency. Custom prompts like captions or specific object categories are pre-encoded and stored as offline vocabulary embeddings. This approach streamlines the detection process without the need for retraining. You can dynamically set these prompts within the model to tailor it to specific detection tasks, as shown below:

    from ultralytics import YOLOWorld

# Initialize a YOLO-World model
model = YOLOWorld("yolov8s-world.pt")

# Define custom classes
model.set_classes(["person", "bus"])

# Execute prediction on an image
results = model.predict("path/to/image.jpg")

# Show results
results[0].show()

  • YOLO-World provides several advantages over traditional Open-Vocabulary detection models:

    • Real-Time Performance: It uses fast CNNs for quick, efficient detection.
    • Efficiency and Low Resource Requirement: YOLO-World maintains high performance while significantly reducing computational and resource demands.
    • Customizable Prompts: The model supports dynamic prompt setting, allowing users to specify custom detection classes without retraining.
    • Benchmark Excellence: It outperforms other open-vocabulary detectors like MDETR and GLIP in both speed and efficiency on standard benchmarks.

  • Train a YOLO-World model on your dataset with the Python API or CLI. Here's how to start training using Python:

    from ultralytics import YOLOWorld

# Load a pretrained YOLOv8s-worldv2 model
model = YOLOWorld("yolov8s-worldv2.pt")

# Train the model on the COCO8 dataset for 100 epochs
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

    Or using CLI:

    yolo train model=yolov8s-worldv2.pt data=coco8.yaml epochs=100 imgsz=640

  • Ultralytics offers multiple pretrained YOLO-World models supporting various tasks and operating modes:

    Model TypePretrained WeightsTasks SupportedTrainingValidationInferenceExport
    YOLOv8s-worldyolov8s-world.ptObject Detection✅✅✅❌
    YOLOv8s-worldv2yolov8s-worldv2.ptObject Detection✅✅✅✅
    YOLOv8m-worldyolov8m-world.ptObject Detection✅✅✅❌
    YOLOv8m-worldv2yolov8m-worldv2.ptObject Detection✅✅✅✅
    YOLOv8l-worldyolov8l-world.ptObject Detection✅✅✅❌
    YOLOv8l-worldv2yolov8l-worldv2.ptObject Detection✅✅✅✅
    YOLOv8x-worldyolov8x-world.ptObject Detection✅✅✅❌
    YOLOv8x-worldv2yolov8x-worldv2.ptObject Detection✅✅✅✅

  • To reproduce the official results from scratch, you need to prepare the datasets and launch the training using the provided code. The training procedure involves creating a data dictionary and running the train method with a custom trainer:

    from ultralytics import YOLOWorld
from ultralytics.models.yolo.world.train_world import WorldTrainerFromScratch

data = {
    "train": {
        "yolo_data": ["Objects365.yaml"],
        "grounding_data": [
            {
                "img_path": "flickr30k/images",
                "json_file": "flickr30k/final_flickr_separateGT_train.json",
            },
            {
                "img_path": "GQA/images",
                "json_file": "GQA/final_mixed_train_no_coco.json",
            },
        ],
    },
    "val": {"yolo_data": ["lvis.yaml"]},
}

model = YOLOWorld("yolov8s-worldv2.yaml")
model.train(data=data, batch=128, epochs=100, trainer=WorldTrainerFromScratch)
Contributors
GLglenn-jocher32RIRizwanMunawar6LALaughing-q6RAraimbekovm4PDpderrenger2FBfbertil1HIhim2him21SHShuaiLYU1Y-Y-T-G1JKjk4e1MAMatthewNoyce1BUBurhan-Q1

Comments244

Ask a question, share your results, or help someone else.

    1. Paula DerrengerPaula DerrengerUltralytics(edited)

      @IamShubhamGupto hey there! 👋

      Absolutely, we're all for community contributions. Training is supported for YOLO-World models in Ultralytics (we recommend yolov8*-worldv2.pt for custom training), see the Train Usage section of the YOLO-World docs:

      from ultralytics import YOLOWorld

model = YOLOWorld("yolov8s-worldv2.pt")
model.train(data="coco8.yaml", epochs=100, imgsz=640)

      If you'd like to add parts of the authors' fine-tuning workflow that aren't covered yet, fork the repo, make your changes, and submit a pull request. Happy coding! 😊👨‍💻

      2
    3. Paula DerrengerPaula DerrengerUltralytics(edited)

      Hey there! 👋

      Yes, you're on the right track! The example you've linked from Roboflow demonstrates filtering detections by area, which is indeed supported by Ultralytics YOLO models. You can achieve similar functionality by manipulating the results object after prediction to filter out detections based on their bounding box area or any other criteria you prefer.

      Here's a quick example of how you might do that:

      from ultralytics import YOLO

# Load a pretrained YOLO26 model
model = YOLO('yolo26n.pt')

# Perform object detection on an image
results = model('path/to/image.jpg')

# Filter detections by bounding box area (e.g., area > 1000 pixels)
filtered_results = [det for det in results[0].boxes.xyxy if (det[2] - det[0]) * (det[3] - det[1]) > 1000]

# Do something with filtered_results

      Hope this helps! If you have more questions or need further assistance, feel free to ask. Happy coding! 😊🚀

      2
    4. sbsbsbsbccccsbsbsbsbccccGitHub

      @IamShubhamGupto嗨呀！👋

      当然，我们全力支持社区贡献和对 YOLO-World 等微调模型的支持。如果你有兴趣做出贡献，那就太好了！🌟

      请注意，虽然我们支持各种 YOLOv8 任务和模式，但yolov8*-world.pt目前不支持 YOLO-World 模型 ( ) 的训练模式。不过，您仍然可以使用预先训练的权重进行推理和验证。

      如果您有与我们现有框架兼容的特定微调流程，我们很乐意看到您的成果！请随意分叉存储库，进行更改，然后提交拉取请求。如果您需要任何帮助或在此过程中有疑问，请随时提出。

      祝你编码愉快！😊👨‍💻

      你好，大佬，我想使用mixgrouding数据集和我自己定义的封闭数据集交叉训练，你有具体的mixgrouding数据集配置教程吗data = { "train": { "yolo_data": ["Objects365.yaml"], "grounding_data": [ { "img_path": "../datasets/GQA/images", "json_file": "../datasets/GQA/final_mixed_train_no_coco.json", }, ], }, }

      model = YOLOWorld("yolov8s-worldv2.yaml")这几个yaml文件哪里可以找到？

  2. meidachenmeidachenGitHub

    Thanks for the great work! I'm wondering which training data has been used for yolov8x-worldv2 model? The model on the official site has "YOLO-Worldv2-X" trained on O365+GoldG+CC3M-Lite, are the training data the same as from here?

    2
    1. Glenn JocherGlenn JocherUltralytics

      @meidachen thank you for your kind words! 😊

      The YOLOv8x-worldv2 model has indeed been trained on a combination of datasets, specifically Objects365, GoldG, and CC3M-Lite. This diverse training data helps the model excel in open-vocabulary detection tasks by leveraging a wide variety of object categories and contexts.

      If you have any further questions or need additional details, feel free to ask. For more information on the YOLO-World model, you can also visit our YOLO-World documentation.

      Happy experimenting! 🚀

  3. DragosjosanDragosjosanGitHub

    The model performance of the Yolo World Hugging Face model (https://huggingface.co/spaces/stevengrove/YOLO-World) is better for my purpose than the standard inference one "YOLOv8x-worldv2" for example.

    Because my lack of domain knowledge regarding this topics I can't make the HF model work on my Mac (not an M processor) so I would like to be able to use the model weights with the abstraction from the ultralytics library.

    Is there any way to use the author's provided weights together with the inference library? For example: model = YOLOWorld('yolow-v8_l_clipv2_frozen_t2iv2_bn_o365_goldg_pretrain.pth')

    From the author's instruction on how to run the model on an image one needs the configuration file and the weights: https://github.com/AILab-CVC/YOLO-World

    The conf file I want to use and the weights are available here: https://huggingface.co/spaces/stevengrove/YOLO-World/blob/main/configs/pretrain/yolo_world_l_t2i_bn_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py https://huggingface.co/spaces/stevengrove/YOLO-World/blob/main/yolow-v8_l_clipv2_frozen_t2iv2_bn_o365_goldg_pretrain.pth

    Thank you in advance :)

    1
    1. Paula DerrengerPaula DerrengerUltralytics(edited)

      Hey there! 😊 Thanks for reaching out with your query.

      Loading the original authors' .pth weights and .py config file into Ultralytics isn't supported: YOLOWorld() accepts Ultralytics .pt weights or .yaml model files only, and the architectures and weight names differ. Use the weights provided by Ultralytics instead, for example yolov8x-worldv2.pt:

      from ultralytics import YOLOWorld

model = YOLOWorld("yolov8x-worldv2.pt")
results = model.predict("path/to/image.jpg")

      Results can differ from the Hugging Face demo because of different weights, preprocessing and settings such as conf and imgsz. If you got any more questions, feel free to ask. 🚀

    2. DragosjosanDragosjosanGitHub

      Hey @pderrenger,

      Thanks your for answering my request! :)

      The "8.1.42" ultralytics Version doesn't have an 'config' instance attribute, so I the provided code snippet raises an error.

      YOLOWorld.__init__() got an unexpected keyword argument 'config'

      Do you suppose I'm doing something wrong?

      Thanks again 👍

    3. Paula DerrengerPaula DerrengerUltralytics

      @Dragosjosan hey there! 😊 It looks like there was a slight misunderstanding regarding model initialization with a 'config' attribute. The YOLOWorld class and the way to load pre-trained models or weights with a specific configuration don't directly support an 'config' attribute as part of the initialization process in the Ultralytics YOLOv8's current implementation.

      For using a model with specific weights, here's the correct approach:

      from ultralytics import YOLO

# Load the model using pre-trained weights
model = YOLO('path/to/your/yolov8s-worldv2.pt')  # Example model

# Now, you can use this model for inference, training, etc.
# Example for inference
results = model.predict('path/to/image.jpg')
results[0].show()

      If you're trying to integrate configurations that don't directly map to the pre-trained models provided by Ultralytics, it would usually involve modifying the model's architecture or training process, which is a bit more involved and typically done within the Python scripts or notebooks rather than through the CLI.

      I hope this clears things up! If you have any more questions or need further help, feel free to ask. 👍

    4. DragosjosanDragosjosanGitHub

      @pderrenger thank you again for your answer.

      Since we don't have the expected model.pt expected by the YOLO class, is there any way to convert the Hugging Face hf_model.pth?

    5. Paula DerrengerPaula DerrengerUltralytics(edited)

      Hey @Dragosjosan! 😊

      Directly loading or converting external .pth weights (such as the Hugging Face ones) into the Ultralytics framework isn't supported, since the model architectures and weight naming conventions differ. Please use the weights provided by Ultralytics, for example yolov8x-worldv2.pt.

      Hope this helps! 🚀

    7. Paula DerrengerPaula DerrengerUltralytics

      Hey there! 🌟

      Thank you for sharing your interest and enthusiasm! 🚀 To directly address your query, here's a concise example showing how to perform inferential tasks using the pre-trained YOLO-World models with our Python API:

      from ultralytics import YOLO

# Load a pre-trained YOLO-World model (e.g., yolov8s-worldv2.pt for object detection)
model = YOLO('yolov8s-worldv2.pt')

# Run inference on an image
results = model.predict('path/to/image.jpg')

# Visualize and/or process the inference results
results[0].show()  # Displays the image with detections

      This snippet loads a YOLO-World model, performs object detection on a specified image, and then visualizes the results. For different tasks or models, simply change the model filename (e.g., yolov8m-worldv2.pt).

      If you're keen on training or further customizing the model, remember that the v2 models (yolov8*-worldv2.pt) support export for deployment. Just ensure you follow the documentation closely for the best outcomes!

      Feel free to reach out for more help or share your progress! Happy coding! 😊

    8. tpsvoetstpsvoetsGitHub

      Hey @Dragosjosan, did you have any success getting the .pth file to work? I am not sure why it is suggested to use a config, while that is currently not supported. And I also don't understand the latest reply by @pderrenger?

    9. DragosjosanDragosjosanGitHub

      Hi @tpsvoets ,

      not with the ultralytics library, no.

      Some of the answers might be auto-generated I guess. Long story short: they don't support it. If you want to convert it to .pt you need to figure it out on your own with pytorch and then also adapt this .pt to the ultralytics configuration.

      Are you also interested in the topic?

    10. junnnnchijunnnnchiGitHub

      Hey @Dragosjosan I am also interested in this topic, did you figure out how to convert .pth to .pt?

    11. DragosjosanDragosjosanGitHub

      Hey @junnnnchi,

      no, sorry. How do you plan on tackling this problem? Maybe we can figure it out together.

    12. tpsvoetstpsvoetsGitHub

      @Dragosjosan ah too bad, it's a shame we cannot get any network file to work with the code. Right now it remains unknown for me how the two demos can differ quite so much. Do we know if the architecture is the same, so if the .pth to .pt conversion works that it will result in something that actually works?

    13. junnnnchijunnnnchiGitHub

      @Dragosjosan I used the roboflow's implementation for YOLO-World. This version gives better results for my project. You might consider this.

    14. DragosjosanDragosjosanGitHub

      @junnnnchi for my application the model from Hugging Face works best... But good that you found something that works. May I ask what your App do? :)

    16. mustfkeskinmustfkeskinGitHub

      How can I integrate the above mentioned files into the ultralytics library. I liked the model results better when working on Huggingface. I noticed that the model size used on Huggingface is a bit bigger size

      The conf file I want to use and the weights are available here: https://huggingface.co/spaces/stevengrove/YOLO-World/blob/main/configs/pretrain/yolo_world_l_t2i_bn_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py https://huggingface.co/spaces/stevengrove/YOLO-World/blob/main/yolow-v8_l_clipv2_frozen_t2iv2_bn_o365_goldg_pretrain.pth

    17. Glenn JocherGlenn JocherUltralytics

      @mustfkeskin unfortunately, integrating the specific .pth weights with the Ultralytics library isn't directly supported. You might need to convert the weights to a compatible format like .pt and adjust configurations manually. For further guidance, consider exploring PyTorch documentation or community forums.

      -1
  4. venuss920venuss920GitHub

    Hi, guys, if we set and fine-tune our own dataset to fine-tune the model, then while fine-tuning the YOLO-World model, it will also learn text prompts while learning object detection as well?

    1. Paula DerrengerPaula DerrengerUltralytics

      Fine-tuning on a standard detection dataset trains YOLO-World to associate visual features with your dataset’s class-name text embeddings; it does not learn arbitrary new prompts as separate annotations or generally train the text encoder. For richer phrase grounding, training requires paired grounding data rather than only YOLO detection labels; see the YOLO-World training reference.

  5. feekyzzfeekyzzGitHub

    为什么会出现单卡训练跑通train 但是多卡会出现类别和grounding之间不匹配的现象？ Why does training run normally on a single GPU, but when using multiple GPUs, there is a mismatch between the nc and the grounding annotations?

    1. Glenn JocherGlenn JocherUltralytics

      多卡下这类 nc 与 grounding 标注不匹配通常更像是 DDP 进程读到了不一致的 data 配置、*.cache 或混合安装环境，而不是单纯类别数设置问题；请先确认你是按 YOLO-World 训练示例 使用 yolov8*-worldv2 和 WorldTrainerFromScratch，然后把完整 traceback、你的多卡启动命令，以及 ultralytics.utils.checks.collect_system_info() 的输出贴在这里，如果最近改过标签或数据集结构，也请先删除 labels 目录下的 *.cache 后再试。

    2. feekyzzfeekyzzGitHub

      感谢您的答复，我按照你的建议，删除了*.cache，确认了完全按照示例使用。

      这是我的train.py

      from ultralytics import YOLOWorld
from ultralytics.models.yolo.world.train_world import WorldTrainerFromScratch

# Option 1: Use Python dictionary
data = dict(
    train=dict(
        yolo_data=["/home/qh1/Qwen/Yw/ultralytics-Yw/datasets/Objects365.yaml"],
        grounding_data=[
            dict(
                img_path="/home/qh1/Qwen/Yw/ultralytics-Yw/datasets/flickr30k/images",
                json_file="/home/qh1/Qwen/Yw/ultralytics-Yw/datasets/flickr30k/final_flickr_separateGT_train.json",
            ),
            dict(
                img_path="/home/qh1/Qwen/Yw/ultralytics-Yw/datasets/gqa/images",
                json_file="/home/qh1/Qwen/Yw/ultralytics-Yw/datasets/gqa/final_mixed_train_no_coco.json",
            ),
        ],
    ),
    val=dict(yolo_data=["/home/qh1/Qwen/Yw/ultralytics-Yw/datasets/lvis.yaml"]),
)
model = YOLOWorld("yolov8-worldv2.yaml")
model.train(
    data=data,
    batch=256,
    epochs=10,
    trainer=WorldTrainerFromScratch,
    workers=16,
    device=[0,2],
    name="Base-v8-worldv2",
)

      这是我的训练脚本.sh

      #!/bin/bash
# Stable RTDETR multi-GPU training script

# 设置使用的物理 GPU
export CUDA_VISIBLE_DEVICES=0,2

# NCCL / DDP 稳定性环境变量
export NCCL_DEBUG=INFO
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
export TORCH_NCCL_BLOCKING_WAIT=1
export TORCH_NCCL_ASYNC_ERROR_HANDLING=1

# 设置随机 master port 避免冲突
PORT=$(shuf -i 25000-30000 -n 1)
echo "Using master port: $PORT"

# 创建日志文件
LOG_FILE="z-train2-02.log"
touch $LOG_FILE

# 后台运行训练
nohup python -u z-train2.py \
    --device 0 2 \
    --master_port $PORT \
    > $LOG_FILE 2>&1 &

echo "Training started, logging to $LOG_FILE"
echo "Use 'tail -f $LOG_FILE' to watch progress"

      这是我的collect_system_info()输出

      Ultralytics 8.3.229 🚀 Python-3.8.20 torch-2.1.0 CUDA:0 (NVIDIA H100 PCIe, 81089MiB)
Setup complete ✅ (128 CPUs, 503.5 GB RAM, 3277.3/24934.1 GB disk)

OS                     Linux-6.8.0-101-generic-x86_64-with-glibc2.17
Environment            Linux
Python                 3.8.20
Install                pip
Path                   /home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics
RAM                    503.51 GB
Disk                   3277.3/24934.1 GB
CPU                    INTEL XEON GOLD 6530
CPU count              128
GPU                    NVIDIA H100 PCIe, 81089MiB
GPU count              8
CUDA                   12.1

numpy                  ✅ 1.24.3>=1.23.0
matplotlib             ✅ 3.7.5>=3.3.0
opencv-python          ✅ 4.13.0.92>=4.6.0
pillow                 ✅ 10.4.0>=7.1.2
pyyaml                 ✅ 6.0.2>=5.3.1
requests               ✅ 2.28.2>=2.23.0
scipy                  ✅ 1.10.1>=1.4.1
torch                  ✅ 2.1.0>=1.8.0
torch                  ✅ 2.1.0!=2.4.0,>=1.8.0; sys_platform == "win32"
torchvision            ✅ 0.16.0>=0.9.0
psutil                 ✅ 7.2.2>=5.8.0
polars                 ✅ 1.8.2>=0.20.0
ultralytics-thop       ✅ 2.0.18>=2.0.18
{'OS': 'Linux-6.8.0-101-generic-x86_64-with-glibc2.17', 'Environment': 'Linux', 'Python': '3.8.20', 'Install': 'pip', 'Path': '/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics', 'RAM': '503.51 GB', 'Disk': '3277.3/24934.1 GB', 'CPU': 'INTEL XEON GOLD 6530', 'CPU count': 128, 'GPU': 'NVIDIA H100 PCIe, 81089MiB', 'GPU count': 8, 'CUDA': '12.1', 'Package Info': {'numpy': '✅ 1.24.3>=1.23.0', 'matplotlib': '✅ 3.7.5>=3.3.0', 'opencv-python': '✅ 4.13.0.92>=4.6.0', 'pillow': '✅ 10.4.0>=7.1.2', 'pyyaml': '✅ 6.0.2>=5.3.1', 'requests': '✅ 2.28.2>=2.23.0', 'scipy': '✅ 1.10.1>=1.4.1', 'torch': '✅ 2.1.0!=2.4.0,>=1.8.0; sys_platform == "win32"', 'torchvision': '✅ 0.16.0>=0.9.0', 'psutil': '✅ 7.2.2>=5.8.0', 'polars': '✅ 1.8.2>=0.20.0', 'ultralytics-thop': '✅ 2.0.18>=2.0.18'}}

      这是我的报错日志

      Traceback (most recent call last):
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/engine/validator.py", line 227, in __call__
    self.loss += model.loss(batch, preds)[1]
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/nn/tasks.py", line 977, in loss
    return self.criterion(preds, batch)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/loss.py", line 463, in __call__
    return self.loss(self.parse_output(preds), batch)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/loss.py", line 468, in loss
    loss, loss_detach = self.get_assigned_targets_and_loss(preds, batch)[1:]
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/loss.py", line 414, in get_assigned_targets_and_loss
    _, target_bboxes, target_scores, fg_mask, target_gt_idx = self.assigner(
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/tal.py", line 98, in forward
    return self._forward(pd_scores, pd_bboxes, anc_points, gt_labels, gt_bboxes, mask_gt)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/tal.py", line 126, in _forward
    mask_pos, align_metric, overlaps = self.get_pos_mask(
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/tal.py", line 164, in get_pos_mask
    align_metric, overlaps = self.get_box_metrics(pd_scores, pd_bboxes, gt_labels, gt_bboxes, mask_in_gts * mask_gt)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/tal.py", line 195, in get_box_metrics
    bbox_scores[mask_gt] = pd_scores[ind[0], :, ind[1]][mask_gt]  # b, max_num_obj, h*w
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.


During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/qh1/.config/Ultralytics/DDP/_temp_jzzt2g3g140562231208544.py", line 13, in <module>
    results = trainer.train()
  File "/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics/engine/trainer.py", line 246, in train
    self._do_train()
  File "/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics/engine/trainer.py", line 520, in _do_train
    self.metrics, self.fitness = self.validate()
  File "/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics/engine/trainer.py", line 752, in validate
    metrics = self.validator(self)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/engine/validator.py", line 227, in __call__
    self.loss += model.loss(batch, preds)[1]
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/ops.py", line 59, in __exit__
    self.dt = self.time() - self.start  # delta-time
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/ultralytics/utils/ops.py", line 69, in time
    torch.cuda.synchronize(self.device)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/cuda/__init__.py", line 783, in synchronize
    return torch._C._cuda_synchronize()
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

[E ProcessGroupNCCL.cpp:915] [Rank 1] NCCL watchdog thread terminated with exception: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Exception raised from c10_cuda_check_implementation at /opt/conda/conda-bld/pytorch_1695392036766/work/c10/cuda/CUDAException.cpp:44 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7a2ad6d91617 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libc10.so)
frame #1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::string const&) + 0x64 (0x7a2ad6d4c98d in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libc10.so)
frame #2: c10::cuda::c10_cuda_check_implementation(int, char const*, char const*, int, bool) + 0x118 (0x7a2ada9939f8 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libc10_cuda.so)
frame #3: c10d::ProcessGroupNCCL::WorkNCCL::finishedGPUExecutionInternal() const + 0x80 (0x7a2a41b46790 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #4: c10d::ProcessGroupNCCL::WorkNCCL::isCompleted() + 0x58 (0x7a2a41b4a5b8 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #5: c10d::ProcessGroupNCCL::workCleanupLoop() + 0x24b (0x7a2a41b60dfb in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #6: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x78 (0x7a2a41b61108 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #7: <unknown function> + 0xdf0e6 (0x7a2ad7c0a0e6 in /home/qh1/Qwen/Yw/yw_env/bin/../lib/libstdc++.so.6)
frame #8: <unknown function> + 0x94ac3 (0x7a2af4294ac3 in /lib/x86_64-linux-gnu/libc.so.6)
frame #9: <unknown function> + 0x1268d0 (0x7a2af43268d0 in /lib/x86_64-linux-gnu/libc.so.6)

terminate called after throwing an instance of 'std::runtime_error'
  what():  [Rank 1] NCCL watchdog thread terminated with exception: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Exception raised from c10_cuda_check_implementation at /opt/conda/conda-bld/pytorch_1695392036766/work/c10/cuda/CUDAException.cpp:44 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7a2ad6d91617 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libc10.so)
frame #1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::string const&) + 0x64 (0x7a2ad6d4c98d in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libc10.so)
frame #2: c10::cuda::c10_cuda_check_implementation(int, char const*, char const*, int, bool) + 0x118 (0x7a2ada9939f8 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libc10_cuda.so)
frame #3: c10d::ProcessGroupNCCL::WorkNCCL::finishedGPUExecutionInternal() const + 0x80 (0x7a2a41b46790 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #4: c10d::ProcessGroupNCCL::WorkNCCL::isCompleted() + 0x58 (0x7a2a41b4a5b8 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #5: c10d::ProcessGroupNCCL::workCleanupLoop() + 0x24b (0x7a2a41b60dfb in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #6: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x78 (0x7a2a41b61108 in /home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
frame #7: <unknown function> + 0xdf0e6 (0x7a2ad7c0a0e6 in /home/qh1/Qwen/Yw/yw_env/bin/../lib/libstdc++.so.6)
frame #8: <unknown function> + 0x94ac3 (0x7a2af4294ac3 in /lib/x86_64-linux-gnu/libc.so.6)
frame #9: <unknown function> + 0x1268d0 (0x7a2af43268d0 in /lib/x86_64-linux-gnu/libc.so.6)

WARNING ⚠️ CUDA OutOfMemoryError in TaskAlignedAssigner, using CPU
[2026-04-21 22:14:56,796] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 2677721 closing signal SIGTERM
[2026-04-21 22:15:02,669] torch.distributed.elastic.multiprocessing.api: [ERROR] failed (exitcode: -6) local_rank: 1 (pid: 2677722) of binary: /home/qh1/Qwen/Yw/yw_env/bin/python
Traceback (most recent call last):
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/distributed/run.py", line 810, in <module>
    main()
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
    return f(*args, **kwargs)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/distributed/run.py", line 806, in main
    run(args)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/distributed/run.py", line 797, in run
    elastic_launch(
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/distributed/launcher/api.py", line 134, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/site-packages/torch/distributed/launcher/api.py", line 264, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
============================================================
/home/qh1/.config/Ultralytics/DDP/_temp_jzzt2g3g140562231208544.py FAILED
------------------------------------------------------------
Failures:
  <NO_OTHER_FAILURES>
------------------------------------------------------------
Root Cause (first observed failure):
[0]:
  time      : 2026-04-21_22:14:56
  host      : fusion-G5500-V7
  rank      : 1 (local_rank: 1)
  exitcode  : -6 (pid: 2677722)
  error_file: <N/A>
  traceback : Signal 6 (SIGABRT) received by PID 2677722
============================================================
Traceback (most recent call last):
  File "z-train2.py", line 24, in <module>
    model.train(
  File "/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics/engine/model.py", line 778, in train
    self.trainer.train()
  File "/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics/engine/trainer.py", line 238, in train
    raise e
  File "/home/qh1/Qwen/Yw/ultralytics-Yw/ultralytics/engine/trainer.py", line 236, in train
    subprocess.run(cmd, check=True)
  File "/home/qh1/Qwen/Yw/yw_env/lib/python3.8/subprocess.py", line 516, in run
    raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['/home/qh1/Qwen/Yw/yw_env/bin/python', '-m', 'torch.distributed.run', '--nproc_per_node', '2', '--master_port', '44421', '/home/qh1/.config/Ultralytics/DDP/_temp_jzzt2g3g140562231208544.py']' returned non-zero exit status 1.
    3. Glenn JocherGlenn JocherUltralytics

      从这份 traceback 看，问题更像是 DDP 子进程混用了本地仓库和 site-packages 的 ultralytics（trainer.py 走的是 /ultralytics-Yw/，但 validator.py、loss.py、tal.py 来自 site-packages），再叠加 WorldTrainerFromScratch 这类自定义组件不适合通过 device=[0,2] 触发自动 DDP；请按 multi-GPU custom code 说明 直接用 torch.distributed.run 启动训练脚本、去掉 model.train() 里的多卡 device 参数，并先统一成单一安装路径（建议只保留一个环境并用 pip install -e .），另外日志里已经出现 TaskAlignedAssigner OOM，先把 batch=256 降下来再复现，这个 OOM 也会连带触发后面的 CUDA device-side assert。

    4. feekyzzfeekyzzGitHub

      感谢您的回复，我按照您的设置已经完全解决了训练的问题。 现在有一个小小的问题，batch不能设置太大，不然仍会在val的过程中触发索引不匹配，这个应该就是你说的问题吧“先把 batch=256 降下来再复现，这个 OOM 也会连带触发后面的 CUDA device-side assert”。

    5. Glenn JocherGlenn JocherUltralytics

      是的，这基本就是同一个问题：val 阶段的 TaskAlignedAssigner 对显存更敏感，大 batch 先触发 OOM 后，后面的“索引不匹配”或 device-side assert 通常只是连锁报错；建议把验证批次设得比训练更小，必要时参考 多卡自定义代码训练说明 先训练、再单独用更小 batch 做 val。

    6. feekyzzfeekyzzGitHub

      你好，为什么在训练了30轮左右，val仍停留在0.019左右

    7. Glenn JocherGlenn JocherUltralytics

      如果你现在是用 WorldTrainerFromScratch 从头训练，这种情况下 30 个 epoch 左右 val 还停在较低水平通常并不异常，尤其是用 LVIS minival 验证时；YOLO-World 从头训练参考 的官方配方本身就是 epochs=100，而 训练常见问题 里也提到从头训练会比加载预训练权重收敛更慢。若到 100 个 epoch 后 precision、recall 和 mAP50-95 仍几乎不变，再把 results.csv、训练曲线和一小份可复现数据贴出来，我们再继续看是否是数据质量或类映射问题。

  6. wcymbnlwcymbnlGitHub

    你好，我在训练自己的数据集的过程中，采用除了yoloworld以外的模型，精度都非常好。可是，一旦利用预训练权重是yolo-worldv2的模型后，一模一样的数据集，map一直为0。训练的代码也是你们提供的，请问如何解决。

    2. Glenn JocherGlenn JocherUltralytics

      @faesray Currently the YOLO-World integration only supports the built‑in CLIP text encoder (ViT-B/32) via build_text_model, so swapping in a custom text encoder is not exposed as a config option—you’d need to modify ultralytics.nn.text_model.build_text_model and retrain, which is a research-level change; for better accuracy with the existing pipeline, use the larger yolov8m/l/x-worldv2 backbones rather than relying on s alone, see the model table in the YOLOWorld reference.

      @yizhihuo99 训练自己数据集时，普通“闭集”微调不需要在 .txt 标签或 yaml 里写整句描述，只要在 data.yaml 里正确填写 names，或在推理时用 model.set_classes([...])，训练器会自动根据类别名生成文本特征；只有像文档示例那样用 WorldTrainerFromScratch 做“检测 + grounding”联合训练时，描述性语言才写在 grounding 的 JSON 里（例如 Flickr30k/GQA 的 caption 字段），并通过 grounding_data 传入，而不是放在 YOLO 的 txt 标签中，具体结构可以参考 WorldTrainerFromScratch 示例.

      @mgh-man 这里的 nc 被硬编码为 80 是刻意设计，它表示“单张图像中最多使用的文本提示数”，而不是数据集的总类别数，所以即使 Objects365 有 365 类，只要一张图里实际出现的不同类别数 ≤ 80，就可以正常训练，不会因为类别 ID > 80 而“无法训练”；目前这个上限只能通过修改 WorldTrainer.get_model 中的 nc=min(self.data["nc"], 80) 以及相应模型代码来变更，暂时不支持通过 yaml 直接把它安全地提高到 365。

      @wcymbnl @mengdxiran 如果同一数据集在其它模型上 mAP 很高，而换成 yolov8*-worldv2 后 mAP 始终为 0，请先确认您是用 YOLOWorld 类在训练，例如：

      from ultralytics import YOLOWorld

model = YOLOWorld("yolov8s-worldv2.pt")
model.train(data="your_dataset.yaml", epochs=100, imgsz=640)

      然后请检查 your_dataset.yaml 中 nc 是否与 names 长度一致、标注类别索引是否从 0 连续编号，并在修正后删除标签目录下的 *.cache 文件再重新训练；如果问题仍在，烦请贴出完整的训练命令、model.train(...) 调用和一次完整的日志，以及 yolo checks 的输出，方便我们进一步定位原因。

  7. mgh-manmgh-manGitHub

    Hello, I reproduced the official results from scratch according to the example and found that objects365 has 365 categories, but I cannot set nc to 365. Model. yaml is set to nc=80 during runtime (model. yaml nc=365 with nc=80), which results in images corresponding to categories with indexes exceeding 80 being unable to be trained. I noticed the following explanation in the code:

    NOTE: This nc here is the max number of different text samples in one image, rather than the actual nc.

    NOTE: Following the official config, nc hard-coded to 80 for now.

    So how can I customize the number of nc?

  8. Yangyang875Yangyang875GitHub

    参考：YOLO-World 模型 - Ultralytics YOLO 文档

    Mr.Y @.***

     

    ------------------ 原始邮件 ------------------ 发件人: "ultralytics/ultralytics" @.>; 发送时间: 2025年9月6日(星期六) 晚上11:04 @.>; @.@.>; 主题: Re: [ultralytics/ultralytics] zh/models/yolo-world/ (Discussion #8458)

    我想问一下，在训练自己的数据集的时候，在哪里添加具有文本描述性的语言，在txt标签文件还是训练时使用的yaml文件

    — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

  9. yizhihuo99yizhihuo99GitHub

    我想问一下，在训练自己的数据集的时候，在哪里添加具有文本描述性的语言，在txt标签文件还是训练时使用的yaml文件

  10. BorhanPetgarBorhanPetgarGitHub

    Hi, thanks for integrating YOLO-World. Is it possible to get the embedding for each bounding box?