Self-hosted image models vs hosted APIs: what actually changes

The short answer: responsibility moves

Choose self-hosting when control over the runtime, model files, or data flow is a concrete requirement and someone can operate the system. Choose a hosted API when you want to integrate generation without maintaining the inference infrastructure. Neither choice establishes better images, lower total cost, or stronger privacy by itself.

The same family can offer both routes. Stability publishes local deployment guidance for Stable Diffusion 3.5 Large and a separate hosted SD 3.5 endpoint. Compare the actual deployment and service, not just the model-family label.

Match the route to a specific constraint

Self-hosting can run on your hardware or rented infrastructure. Renting a GPU does not remove the need to evaluate the infrastructure provider and your own configuration.

A downloadable model still has a license

Stability's license page lists the SD 3.5 suite among covered Community models and describes Community eligibility below USD 1 million in annual revenue, with Enterprise licensing for covered commercial use above that threshold. This concerns model use, not a free allowance of hosted API calls. Check the actual agreement and your organization's circumstances rather than applying one summary to every Stable Diffusion release.

Even close relatives differ: Black Forest Labs publishes FLUX.2 Klein 4B under Apache 2.0, while the 9B weights carry its Non-Commercial License. Permission to use a hosted endpoint is a separate question from permission to deploy weights. Start with the FLUX license notes, then read the exact variant's license.

Privacy follows the whole data path

A self-hosted system can be designed to keep inference inputs inside a chosen environment, but inspect storage, logs, extensions, telemetry, backups, and administrator access as well. A local interface that calls a remote model still transmits the submitted material. Document where every component runs before calling the workflow private.

Hosted policies also differ. OpenAI's API data-controls table lists no training use for image generation/editing endpoints, with default abuse-monitoring retention of 30 days and additional conditions for Zero Data Retention. Stability's privacy policy permits training-related uses of personal data and describes an opt-out. These are scoped provider statements, not interchangeable promises.

Review the endpoint, account configuration, applicable contract, and any intermediary. The GPT Image privacy record distinguishes image endpoints from other API workflows.

A base model is only part of a production workflow

The hosted SD 3.5 reference documents text-to-image and image-to-image generation. Stability separately lists editing, control, and upscaling services on its API service page. Downloading a base model does not establish that you have an equivalent implementation of each surrounding service.

Write down the entire job: input validation, generation, masked correction, enlargement, export, and storage. For each step, identify the component and supported format. Use the Stable Diffusion capability notes to avoid silently attributing a service feature to every model.

Control helps reproducibility, but a seed is not enough

Self-hosting lets you retain a chosen checkpoint and software configuration. Record the model revision, pipeline version, scheduler, seed, dimensions, precision, and relevant hardware. Hugging Face's Diffusers reproducibility guide explains randomness controls and deterministic algorithms, including potential performance trade-offs.

Do not promise identical output across changed environments from a seed alone. Hosted services can also change behavior: Stability documents the retirement of SD 3.0 APIs and rerouting to 3.5. Keep representative acceptance examples and rerun them when a dependency or endpoint changes.

Compare total operating cost, not weights versus a call

For self-hosting, include hardware or rented compute, idle capacity, storage, maintenance, monitoring, and staff time. For hosted access, include requests, edits, rejected attempts where billable, storage, and integration work. Stability's hosted service is credit-metered; the model license answers a different cost question.

Use your expected workload rather than an invented break-even number. A sporadic workflow and a continuously busy batch system have different economics. Track cost per accepted deliverable and correction time separately. Compare both routes against the same workload and operating requirements.

Assign the operational work and test an exit path

Before choosing self-hosting, name who handles capacity, failed jobs, updates, access controls, backups, and content safeguards. Stability's model card specifically calls on deployers to implement safeguards for their application. Hosted inference shifts some work to the provider, but your application still needs error handling, spending controls, and a plan for unavailable or changed endpoints.

For a small evaluation, use representative inputs you are allowed to process and run the full delivery workflow. Check one retry, one revision, and one export. Preserve prompts, settings, original references, and accepted files in your own records. A hybrid approach is possible, but verify the data destination and output differences before using a second route as a fallback.

Sources and scope

Official sources were reviewed on October 21, 2026. Recheck the exact license, service terms, and endpoint behavior before implementation.