Skip to main content
Hand-drawing every annotation stops making sense the moment you have a model that gets 80 percent of them right. The Models page (under Flow in the top navigation) registers your inference endpoints so PixlHub can call them for pre-labeling, and humans spend their time correcting instead of drawing.
PixlHub Models page for configuring ML models for automated pre-labeling

The Models page: registered models on the left tab, every execution on the Run History tab.

​
Bring your own model

PixlHub treats a model as an HTTP endpoint you own: your YOLOv8 service, your SageMaker successor, your internal detector behind an API. Registration is one form:
PixlHub Add Model dialog with provider, model type, endpoint URL, API key, input config and output field mapping

Add Model: endpoint, credentials, and the JSON contracts that adapt any API shape to PixlHub.

The two JSON contracts are the point: instead of forcing your API into a fixed schema, you describe the request and map the response, and almost any detector fits without a wrapper service.

​
Where registered models act

The AI Tools section of the annotation palette exposes model assistance during labeling: propose annotations on the current task, then correct by hand. See AI-assisted annotation for the editor-side loop.

​
Run History

Every execution lands in the Run History tab: which model, which scope, when, and how it went. Treat it as the audit trail for automated labeling; when a batch of pre-labels looks off, the history tells you which model version produced it.

​
Pre-labeling that actually saves time

  • Correcting beats drawing only when the model is decent. Below roughly usable accuracy, deleting wrong boxes costs more than drawing; pilot on a small batch first.
  • Keep humans in the loop. Pre-labels are proposals, not truth: run them through the same review gate as human work, and watch the Reviewer Edit Rate.
  • Version your endpoint names. “YOLOv8-shelf-v3” in the name field makes Run History legible six months later.

​
Frequently asked questions

No. Models are an accelerator, not a requirement; the editor’s manual tools and the smart palette work without any registered model.
On your endpoint, on your infrastructure. PixlHub sends requests and maps responses; the model, its weights and its compute remain entirely yours.
Language-model workers are configured separately as Agents, with providers like OpenAI, Anthropic, Google, Grok and local Ollama.

​
Next steps

Agents

LLM workers for text, NER and captioning workflows.

AI-assisted annotation

The editor-side experience of model assistance.