Enhance AI with Your Content: A Step-by-Step Guide

Capture Wallet powered by Numbers Protocol plays a crucial role in building ethical AI with user-owned and user-controlled dataset. Combining AssetTree (metadata) and Nid (decentralized content storage) together, digital content (image, video and music files) in the Capture wallet forms a robust training dataset for AI. This setup allows AI models to access user content securely, with explicit wallet permission and AI training consent.

By adhering to ERC-7517, it filters content that allows AI training using miningPreference field in the AssetTree , further protecting the creators' rights and ensuring ethical AI practices.

The page demonstrates the use case that developers can utilize Capture, HuggingFace, and Colab to train your own models with your registered assets and generate new AI inference images. By following these steps, you can leverage our service to enhance your creative capabilities and produce high-quality images. By filtering content following ERC-7517,

Access Colab with Hugging Face Diffusers

• Click on the provided Colab link to open it in your web browser.

• Navigate to the urls in the Settings for teaching your new concept section.

Before

Access Data From Capture Wallet

• Replace the entire code block labeled 'urls' with the provided code. The code will list links to the creator's images that match the configured mining preference and AI training tag.

import requests


def list_creator_images_nid(capture_token, your_capture_email):
 resp = requests.get(
        "https://api.numbersprotocol.io/api/v3/assets/",
        params={
            "owned_only": False,
            "is_original_owner": True,
            "owner": creator_email,
        },
        headers={"Authorization": f"token {capture_token}"},
        timeout=30,
    )
    resp.raise_for_status()
    assets = resp.json()["results"]
    output = []
    for asset in assets:
        if not asset["asset_file_mime_type"].startswith("image/"):
            continue
        output.append(asset['id'])
    return output

def filter_matched_urls(
    nids,
    capture_token,
    ipfs_gateway,
    matched_mining_preference_key,
    matched_mining_preference_value,
    matched_tag_key,
    matched_tag_value,
):
    output = []
    for nid in nids:
        resp = requests.get(
            "https://eohakmlw1xjl7ut.m.pipedream.net",
            params={"nid": nid},
            headers={"Authorization": f"token {capture_token}"},
            timeout=30,
        )
        resp.raise_for_status()
        commits = resp.json()["commits"]
        if not commits:
            continue
        payload = []
        for commit in commits:
            payload.append({"assetTreeCid": commit["assetTreeCid"], "timestampCreated": commit["timestampCreated"]})

        resp = requests.post(
            "https://eoqlryxobmjf9kw.m.pipedream.net",
            json=payload,
            headers={"Authorization": f"token {capture_token}"},
            timeout=30,
        )
        resp.raise_for_status()
        assetTree = resp.json()["assetTree"]
        if assetTree.get(matched_mining_preference_key) != matched_mining_preference_value:
            continue
        if assetTree.get(matched_tag_key) != matched_tag_value:
            continue

        output.append(f'https://{ipfs_gateway}/ipfs/{nid}')
    return output


# @markdown Use your [`capture_token`](https://docs.captureapp.xyz/capture-account-and-wallet) to set the URLs
# @markdown to the images of the concept. These images belong to the `creator_email` and are allowed for AI training.
# @markdown They should have a matched tag equal to `matched_tag_value`
capture_token = ""  # @param {type:"string"}
creator_email = "[email protected]"  # @param {type:"string"}
ipfs_gateway = "ipfs-pin.numbersprotocol.io"  # @param {type:"string"}
matched_mining_preference_key = "miningPreference"  # @param {type:"string"}
matched_mining_preference_value = "aiTrainingWithAuthorship"  # @param {type:"string"}
matched_tag_key = "aiTrainingtag"  # @param {type:"string"}
matched_tag_value = "Elephant"  # @param {type:"string"}
urls = filter_matched_urls(
    list_creator_images_nid(capture_token, creator_email),
    capture_token,
    ipfs_gateway,
    matched_mining_preference_key,
    matched_mining_preference_value,
    matched_tag_key,
    matched_tag_value,
)

After

• Make sure to complete the fields: fill in capture_token with your Capture token, creator_email with your Capture email, and matched_tag_value with the tag assigned to assets that will be used in this training.

Configure Training Parameters

• Configure the instance_prompt in the Settings for your newly created concept subsection.

• Navigate to the Run the code with your newly trained model section. Ensure to tick the save_concept option, specify name_of_your_concept, choose where_to_save_concept, and input hf_token_write with your Hugging Face token in the Save your newly created concept? subsection.

• Click on Show code and replace the code api = HfApi() with api = HfApi(token=hf_token_write) in the same section.

Train Your Model

• Run each subsection in the Initial setup, Settings for teaching your new concept, and Teach the model the new concept (fine-tuning with Dreambooth) sections to train your model.

Improve training data (Optional)

• Dreambooth automatically crops center square images. If your target isn't centered but you want to refine the training data, here's what to do after you've completed the Setup and check the images you have just added subsection but before moving on to Settings for your newly created concept:

  • Download the images from Colab to your local machine.

  • Remove the original images from Colab.

  

  • Manually crop the images as necessary and save them locally.

  • Once cropped, upload the newly saved images back to Colab.

  

Save Your Model

• Run the Save your newly created concept? subsection in the Run the code with your newly trained model section to save your trained model.

Generating AI Inference Images

• Click the link Click here to access it to open the Hugging Face model.

• Run the Text-to-Image task to generate a new image.

Examples

• Usopp toy

creator_email: [email protected]
matched_tag_value: Usopp
instance_prompt: Usopp toy
Text-to-Image: The Usopp toy is riding a bicycle

Concept Images

Inference Images

• Elephant toy

creator_email: [email protected]
matched_tag_value: Elephant
instance_prompt: elephant toy
Text-to-Image: oil painting of elephant toy in style of van gogh

Concept Images

Inference Images

• Lion toy

creator_email: [email protected]
matched_tag_value: Lion
instance_prompt: sks toy
Text-to-Image: A cat doll resembles a sks toy

Concept Images

Inference Images

• Usopp toy (Improve training data)

creator_email: [email protected]
matched_tag_value: Usopp
instance_prompt: Usopp toy
Text-to-Image: The Usopp toy is riding a bicycle

Concept Images

Inference Images