Generate one image with Stable Diffusion 3.5 Large in Replicate’s web Playground and save it as a file. This exercise uses a named model on a hosted service, so you can evaluate the result without installing a checkpoint or configuring your own GPU.

Before you start

Open the Stable Diffusion 3.5 Large Playground and sign in to a Replicate account with permission to run predictions. Check the per-output charge on the model page and the billing settings before running. Use a fictional subject with no uploads. Replicate’s web predictions retain data under the host’s policy, which differs from its API policy. Replicate’s data retention policy

How to use Stable Diffusion: your first result

Step 1. Confirm the exact model

Check that the model page identifies stability-ai/stable-diffusion-3.5-large. Open the Playground’s form view. Other checkpoints with “Stable Diffusion” in the name can offer different controls and terms. This exercise uses 3.5 Large throughout.

Step 2. Fill the prompt and output settings

Remove any demo text and paste the example in prompt. Leave the input image empty for a text-only generation. Select square aspect_ratio and set output_format to the format you need, such as the default WebP. Keep the other settings in place for the first image.

Step 3. Run the form once

Start the prediction and wait for the displayed status to finish. Check the result from your own run rather than the sample already shown on the model page. Record the model slug, the Playground, prompt, ratio, format and any seed shown. If a run fails, inspect its status and billing before submitting it again.

Step 4. Inspect and save the result

Open the output and check the mug’s handle, rim and contact shadow. Make sure there is one mug, no lettering and enough room around it for a square crop. Use the output Download action and reopen the file locally to check the format and dimensions.

Step 5. Check the model and host terms

Keep the 3.5 Large model card with your notes. It identifies the exact checkpoint and its license. Check the relevant Stability AI terms for how you plan to use the model or output and Replicate’s terms for the hosted service. The hosting charge applies separately from model licensing.

Example input and expected result

A single white ceramic mug with one visible handle, centered on a light oak table against a pale gray wall, soft daylight from the left, realistic contact shadow, square product photograph, plain surface, no lettering or logos.

The expected result is one usable square image of a mug and a saved file. Inspect the inside of the handle for broken edges, the top rim for extra lines and the base for a shadow that connects the cup to the table. Record one concrete defect if the result needs a second pass.

Three tips for better results

  1. Keep the exact model name attached to the saved prompt.
  2. Start from a short brief with one subject and one lighting direction.
  3. Compare a second run only after identifying the first image’s specific failure.

Common beginner problems

The Playground asks for billing

Review your organization’s billing permissions and the displayed model price. Confirm the charge before authorizing a prediction.

The aspect ratio control has no effect

Clear the input image. The model form documents that the aspect-ratio field is ignored when an input image is supplied.

What to try next

  1. Revise a single constraint and compare the next image.
  2. Practice writing a focused one-subject image brief.
  3. Use the Stable Diffusion overview to compare model licenses, hosted services and local setup.

Frequently asked questions

Do I need to install software or create an API key?

The Playground runs in your browser with a Replicate account. Use the API only for a separate software integration.

Does the model license make hosted runs free?

Replicate charges for running the hosted model. The Stability AI license governs use of the checkpoint under its terms. Stability AI licensing

Can I use another checkpoint with the same instructions?

Check its own model card and Playground fields. Settings and rights are specific to the exact checkpoint and interface.