--- tags: - audio - trap - hip-hop - music - text-to-audio - generative-audio - stable-audio - stable-audio-open - finetune language: - en base_model: stabilityai/stable-audio-open-1.0 base_model_relation: finetune pipeline_tag: text-to-audio license: other license_name: stabilityai-community-license license_link: https://stability.ai/license datasets: - custom model_name: Gluten – Trap, Hip-Hop & Pop Finetune --- # ![image](https://cdn-uploads.huggingface.co/production/uploads/67e48dcf55a1ff318bab67a8/KuSk2epliP6y_ZgxC7Wi6.png) A fine-tuned version of **Stable Audio Open** focused on generating **drumless, short musical loops, which seemlessly loop**, **drum-loops** and **one-shot sample-style clips**, with strong conditioning on **BPM**, **key**, and **mood** metadata. It is the base-version of the sample-generation model used in [Kurt](https://trykurt.com) > **Best at:** creating sample-ready loops in **Trap / Hip-Hop / Pop** (especially melodic trap-style loops). > **Conditioning:** BPM-aligned, key-aligned, mood-aligned *to the extent the training labels and model generalization allow*. ## Model Details - **Model type:** Text-conditioned generative audio model (fine-tune of Stable Audio Open) - **Base model:** `stabilityai/stable-audio-open-1.0` - **Domain:** Music loops (Trap / Hip-Hop / Pop) - **Primary use:** Generating loopable musical ideas and samples - **Languages:** English prompt metadata (structured) ### Prompt / Conditioning Format Use the following structured format (recommended): `Format: Solo | Genre: Trap | Sub-Genre: Melodic Trap | Instruments: Piano, Synth Pad | Moods: Melancholic, Reflective | Styles: Catchy, Smooth | Tempo: Medium | BPM: 135 | Key: Dm` **Make sure to calculate the time-duration in seconds according to your input-BPM and 4/4-beat for best results ie. seconds = (bars * 240) / BPM** **Field guidance** - **Format:** e.g., `Solo`, `Full`, `Loop`, etc. (use what matches your dataset convention) - **Genre/Sub-Genre:** e.g., Trap / Melodic Trap, Hip-Hop / Boom Bap, Pop / Electropop - **Instruments:** comma-separated, e.g., `Piano, 808, Synth Lead` - **Moods/Styles:** comma-separated descriptors - **Tempo:** e.g., `Slow`, `Medium`, `Fast` (even if BPM is provided this can help to inforce the mood) - **BPM:** integer (e.g., 135) - **Key:** e.g., `Dm`, `F#min`, `Cmaj` - **Length:** seconds = (bars * 240) / BPM ### Example Prompts All created with CFG = 9 - 'Format: Solo | Genre: Trap | Sub-Genre: Trap | Instruments: Synth Pad, Synth Lead | Moods: Epic, Dark | Styles: Atmospheric, Building | Tempo: Mid | BPM: 140 | Key: Em,'

--- ## Training ### Training Setup - **Hardware:** 8× NVIDIA RTX 4090 - **Epochs:** 20 - **Objective:** Fine-tune Stable Audio Open to better follow structured metadata for loop generation in Trap / Hip-Hop / Pop. ### Data - **Dataset size:** ~80,000 loop clips - **Content:** Trap, Hip-Hop, Pop loops (music-only or drums-only) - **Strength:** high consistency in short musical phrases / loop structure #### Labeling / Metadata Labeling and metadata generation were derived using: - **Qwen Omni** (for multi-modal/semantic labeling) - **MuQ / MuLan-based audio-text embedding alignment** (for music semantic tags) - These labels were used to create structured conditioning fields (genre, sub-genre, instruments, mood/style descriptors, BPM, key). --- ## Intended Use ### Primary Intended Use Cases - Producing **loopable musical ideas** for Trap/Hip-Hop/Pop - Generating **sample packs / sample candidates** - Rapid prototyping of **melodic loops**, chord beds, simple motifs, textures ### Out-of-Scope / Not Recommended - Full-length song generation (beyond the clip duration supported by the base model/tools) - Vocals/lyrical content reliability --- ## Limitations - **BPM/Key compliance is not guaranteed:** Even with conditioning, outputs can drift. Please ensure the prompted length aligns with the BPM and bars. - **Genre bias:** Tuned for Trap/Hip-Hop/Pop loops; other genres may sound less consistent. - **Prompt sensitivity:** Best results come from the provided structured format. Use a LLM to convert natural language prompts to this format. --- ## Recommendations for Best Results - Keep prompts **structured** and **specific** (instruments + moods + BPM + key). - If BPM/key are critical: - Generate multiple candidates and select the closest match. - Use light post-processing for perfect grid/key lock. - Use “Format: Solo” everytime --- ### Stability AI Community License This model is a fine-tuned version of the Stable Audio Open 1.0 model by Stability AI. It is licensed under the Stability AI Community License, which allows: Free use for research and non-commercial purposes. Limited Commercial use for entities with annual revenues below USD $1M. An enterprise license is required for commercial use by entities with annual revenues exceeding USD $1M. For detailed terms, refer to the original LICENSE. Additional Terms for Fine-Tuned Model As per Stability AI, as a fine-tuner, no upfront payment is required under this license. However, it is crucial that if this model is used in any commercial application in which annual revenue exceeds USD $1M, you must contact Stability AI for enterprise licensing. Further details can be found here. }