Infrastructure and model catalog
Replicate packages models as runnable software with inputs and outputs, then exposes them through a web playground and API. Official models are maintained with stable interfaces and predictable metering; community models remain maintained by their publishers. Users can also create private models. This makes Replicate part catalog, part GPU inference layer, and part deployment service. A public listing is not proof that Replicate authored the weights or guarantees every community model's documentation.
ComfyUI, diffusion, and fine-tuning
The official ComfyUI guide converts a ComfyUI graph into API JSON and runs it through a hosted model. It documents SDXL, Stable Diffusion checkpoints, ControlNet, AnimateDiff, Stable Video Diffusion, LoRAs, and custom nodes. Separate documentation covers fine-tuning FLUX.1 with training images and running the result through an API. These are hosted versions of workflows that could otherwise use local interfaces and GPUs; network upload, cold start, billing, and service policy become part of the pipeline.
Custom models and deployments
Custom models are packaged with Cog, which defines dependencies and prediction code and generates an API server. A deployment adds chosen GPU hardware, minimum and maximum instances, autoscaling, monitoring, private endpoint control, and rolling updates. This is useful when a creator needs repeatable inference rather than a manual notebook, but it also makes container security, dependency provenance, model license, and cost configuration the user's responsibility. Keeping an instance warm can reduce cold starts while increasing idle cost.
Policy, privacy, and reproducibility limits
Replicate is not an adult platform, and technical support for custom code or open models does not establish permission for adult or likeness-based workloads. Platform policies, model-specific licenses, consent, and lawful processing still govern a deployment. Inputs, outputs, training data, and logs pass through hosted systems, and public predictions or model examples may have different visibility from private models. Model versions also matter for reproducibility: a mutable community model, custom dependency change, or new deployment version can alter outputs even when the prompt remains the same.
Technical context
Replicate documents ComfyUI deployment, FLUX fine-tuning and model APIs; the linked dossiers explain the model and workflow layers beneath its hosted service.