stabilityai / stable-audio-3-medium

Overview

Stable Audio 3 Medium is an audio generation model from Stability AI. It uses latent diffusion over a compressed audio representation from a semantic-acoustic autoencoder. (huggingface.co)

You write a text description and it produces a stereo music clip in a chosen format like MP3 or WAV. This is useful when you need custom music beds, loops, or draft tracks without clearing a sample library first. (fal.ai)

What you can build

• Background music for videos, trailers, and reels
• Game music drafts for menus, levels, and ambient zones
• Podcast intro and outro music sketches
• Mood boards for composers and audio directors
• Variations on a musical idea by changing the prompt and seed

Inputs

• A required text description of the music you want (English works best). Include style, instruments, mood, and BPM when you can. (fal.ai)
• An optional text description of what to avoid in the result (for example, “no vocals” or “avoid harsh distortion”). (fal.ai)
• The target length in seconds. The Medium model supports up to 380 seconds (about 6 minutes 20 seconds). (fal.ai)
• The number of denoising steps to use during sampling. More steps can change detail, but the post-trained checkpoint is designed to look good at low step counts. (fal.ai)
• A guidance strength value that controls how tightly the output follows your prompt. (This control matters most on base, non-distilled checkpoints.) (fal.ai)
• An optional seed number for repeatable results. Omit it for a random result each run. (fal.ai)
• An optional toggle to expand your prompt with an LLM, which can help when your prompt is short. (fal.ai)
• An optional toggle to enable an NSFW safety checker. (fal.ai)
• Your preferred audio container format for the output (MP3 is the default, and WAV/FLAC are available for lossless exports). (fal.ai)
• For compressed formats, an output bitrate string (for example, values like 192k or 320k). (fal.ai)
• An optional “inline return” mode that can return audio as a data URI instead of a hosted download. (fal.ai)

Outputs

The model returns a generated audio file plus metadata.

• An audio file object that includes a downloadable URL and file metadata such as MIME type, filename, and size. (fal.ai)
• The seed that was used for the generation, so you can reproduce or iterate. (fal.ai)
• The prompt text that was actually used for the generation. (fal.ai)

Recommended settings

• Use 8 denoising steps for the post-trained Medium checkpoint. It is tuned to look good at that default. (fal.ai)
• Keep guidance at 1.0 unless you are using a base (non-distilled) checkpoint. Guidance has limited effect on distilled checkpoints. (fal.ai)
• Turn on prompt expansion when your prompt is short or vague. Leave it off when your prompt is already detailed. (fal.ai)
• Export MP3 for quick review. Export WAV or FLAC when you plan to edit or master the result. (fal.ai)

Limitations

• Maximum generation length for the Medium model is 380 seconds. Requests longer than that may fail or be truncated. (fal.ai)
• The model is aimed at music generation. For pure sound effects, Stable Audio 3 Small SFX may be a better fit. (github.com)
• Don’t expect clean, intelligible vocals. It can produce vocal-like textures, but words typically won’t be understandable. (github.com)
• Vague or conflicting prompts often produce generic results. Add genre, instrumentation, and tempo to improve control. (github.com)
• If you export to lossy formats (like MP3), compression artifacts can appear in cymbals and dense mixes. (fal.ai)

Safety & compliance

• You can enable an NSFW safety checker. Turn it on if you accept public or user-generated prompts. (fal.ai)
• Stability AI states Stable Audio 3.0 models are trained on fully licensed data, and the open weights for Small and Medium are provided under a Stability AI Community License. (stability.ai)
• Under Stability AI’s licensing terms, commercial use depends on your organization’s revenue tier. Review the Stability AI license terms before shipping a product. (stability.ai)