fal Playground Guide: Inputs, Seeds and Settings

Note: fal is a developer API platform, so Dreamdrive doesn’t capture from it automatically. The advice below still applies; download your outputs and import them into your library to keep their prompts and settings with them.

fal (fal.ai) is a hosted inference platform: fast image and video models served behind an API, with a web playground for every model so you can try it without writing code. You pay per use rather than by subscription. This guide walks through the playground form, explains the inputs most models share, and shows how to turn a lucky result into a repeatable one.

What the fal playground is

Every model on fal.ai has its own page. The gallery lists Flux variants for images, and video models such as Kling, Hailuo (MiniMax), Wan, LTX and Veo at various times; the line-up changes often, so what you see today may differ. A model page has a playground tab with a form on one side and the output on the other, an API tab with code snippets, and usually a short description with example results.

You need an account and a payment method. Billing is per request, and the cost depends on the model, the output size or the number of seconds of video. Check the platform’s current pricing page rather than relying on figures from older posts.

Because each model exposes its own inputs, the form is never quite the same twice. The field names below are the common ones; whether a given field exists, and what it is called, depends on the model you have open.

Inputs you will see on most models

• Prompt. The text description. Image models want subject, setting, light and style; video models want motion and camera.
• Negative prompt. Only on models that support it. Many Flux endpoints do not.
• Image URL or upload. Used by image-to-image, editing and image-to-video models. When you upload a file, fal hosts it and the URL appears in your request, which is why the request JSON can be replayed later.
• Image size or aspect ratio. Either named presets (square, portrait, landscape) or explicit width and height.
• Number of images. How many variations one run produces. Each one is billed.
• Seed. Random unless you set it. Same seed, same inputs, same model gives the same result.
• Steps (often num_inference_steps). More steps cost more time and money with diminishing returns.
• Guidance scale. How strictly the model follows the prompt. Too high looks over-sharpened and literal; too low wanders.
• Safety checker and output format. Keep the checker on for client work. PNG for stills you will edit, JPEG or WebP for web.
• Duration and resolution on video models, sometimes with camera or motion fields.

Step by step: a repeatable run

1. Pick a model. For a first test choose a fast Flux text-to-image endpoint; it is cheap and returns in seconds.
2. Write the prompt and leave every other field at its default. You want to know what the model does before you steer it.
3. Run it. Note how long it took and the cost indicator if the page shows one.
4. Open the request details. Beside the output you will find the inputs as JSON, including the seed that was used. Copy that block.
5. Fix the seed and change one input. Paste the seed back in, raise or lower guidance, run again and compare side by side. One change per run is the only way to learn what each control does.
6. Save the output and the request. Results stay on the request page for a while, but not forever. Download the file and keep the JSON next to it.

Settings at a glance

From playground to API

Once a run works, switch to the API tab. It shows the same request as code in several languages, and every field in the form maps to a key in that JSON. Set things up visually, then copy the call. Larger jobs go through a queue, so you submit a request, get an ID, and fetch the result when it is ready. Keep the request ID with your files; it is the only link back to the exact inputs if you did not copy them.

If you are weighing fal against a similar service, see Replicate vs fal for creatives. For video endpoints specifically, read the fal video models guide.

Common problems

• An image URL fails. The link must be publicly reachable. Upload the file instead and let fal host it.
• The same prompt gives a different picture. The seed was random, or the endpoint was updated. Record the full model identifier and the seed with every result you want to keep.
• A run is slow. Steps are high, the resolution is large or the queue is busy. Test small and scale up at the end.
• The bill is higher than expected. Resolution and number of images multiply the cost. Generate one image at a time until the prompt is right.

Try it

Model: a Flux text-to-image endpoint
Prompt: product photo of a matte black ceramic teapot on a slate surface, soft window light from the left, shallow depth of field, plain grey background
Image size: landscape 4:3
Number of images: 1
Seed: 42
Steps, guidance: model defaults

Run it once, then keep the seed at 42 and change only the guidance scale up and down. Compare how literal the teapot becomes before you touch the prompt again.