Choosing a model
Start with the result you need and the inputs you have. The model catalog lists configurations; the workflow guides explain specific graphs and remaining manual steps. A catalog entry is not a promise that a particular prompt will produce an acceptable result.
Match the model to the job
Text and chat
Bring: Representative prompts, the response format you need, and an estimate of conversation length.
Check: Choose the exact model variant. Check memory requirements and whether your application needs a dedicated workspace or an existing serverless text endpoint.
Images and edits
Bring: A description of the output; for editing, the source image and any identity or style references.
Check: An image-generation model and an image-edit model solve different tasks. Confirm the graph’s model files, supported inputs, resolution, and any LoRAs before launching.
Video
Bring: The required first frame, reference image, or driver video, plus a motion prompt and target format.
Check: Read the exact template’s instructions. Some require importing a separate workflow JSON. Frame count, canvas size, model variant, and motion complexity affect time and output quality.
Speech
Bring: The script, desired language, and a voice choice or suitable reference recording when the model supports it.
Check: Check the selected model’s supported voice and language inputs. Use the speech API for its supported request format, or a dedicated interface when you need its additional controls.
What a dedicated deployment buys
You rent a running GPU workspace with the selected runtime and configuration. ModelPilot prepares supported files and shows startup, logs, and lifecycle controls. You supply the required inputs, run the workflow, review its output, and download what you need. Templates can still require manual setup inside the workspace.
Review memory, time, and cost
- A GPU recommendation starts from the configured model requirements. Higher resolution, longer clips, larger batches, or extra models can need more memory.
- Use the pricing page to compare listed rates, then confirm the exact instance and rate in the setup review. Recheck any fallback offer.
- Startup depends on capacity, container caching, model downloads, and custom nodes. Running GPU time can be billable while the workspace is still preparing.
- An idle dedicated GPU continues to cost money. Stop compute when finished; download important files and delete resources you no longer need, since stopped storage can remain billable.
When to use an API
The image and speech APIssupport a narrower set of request formats and configured models. Chat requests route to a text model already deployed in your account or an existing owned serverless endpoint. A model being listed in the dedicated catalog does not make it available through every API.
Before the first paid run
- Pick the exact model or workflow and read its limitations and required inputs.
- For a custom graph, use the workflow checker to identify recognized dependencies and unresolved files. Analysis does not execute the graph.
- Define one output you can judge, such as a short passage, one image, or one clip. Keep enough budget for startup and review before expanding the batch.
- Follow the setup guide and keep the interface instructions available while you work.