The paper, in the repo
The official version (v6, Chinese and English, PDF + Word) ships in paper/ — the single citable version, with a CITATION.cff alongside.
Research · Paper & Open Materials
Botanica-OS
Organic operations on neural latent spaces — interaction paradigms and aesthetic boundaries of AI-assisted Botanica/IDM timbre design.
License GPL-3.0 (code) · CC-BY 4.0 (prose)Paper v6 · CN / EN · PDF + WordStudy N = 12 · mixed methodsStack Python · PyTorch · React 18
Features
What it does
The open-materials repository accompanying the paper. It proposes the framework of organic operations, which reorganizes neural latent-space interaction into three botanically-metaphor paradigms, instantiates them in a prototype system built on a unified differentiable-DSP real-time pipeline, and quantifies the differences in a mixed-methods user study with N = 12 Botanica/IDM producers and ML researchers. The official paper (v6, Chinese and English, PDF + Word) ships in paper/.
Three organic operations
Graft — a continuous transition between two timbre parents; Hybridize — intention-guided variation around a CLAP semantic point; Cultivate — bandit-style trajectory optimization driven by preference feedback.
Mixed-methods user study
N = 12 producers and ML researchers: Graft tops overall SUS (78.3), Cultivate tops agency (5.9/7), and Hybridize's experienced controllability falls below its conceptual placement — one of the paper's core findings. Task protocols, questionnaires and interview guides included.
Reference implementation
Eqs. (2)–(4) and Algorithm 1 reimplemented with numpy only, unit-tested, in reference/.
47 ms real-time pipeline
End-to-end latency 47 ± 3 ms on CPU inference (16 kHz, 128-sample blocks), meeting the <100 ms performative-interaction requirement (DR5).
Full reproduction kit
Latency breakdown, summary of results (Table 7), baseline fairness controls, hyperparameters (Table A1), the seeds.yaml random-seed registry and the re-analysis script.
Details
The three organic operations
| Paradigm | Input | Time structure | Formula | Description |
|---|---|---|---|---|
| Graft | Dual parents z_a, z_b + mixing coefficient α | Single-step, linear | Eq. (2) | A continuous transition between two timbre parents |
| Hybridize | Semantic condition vector c (CLAP text embedding) | Single-step, probabilistic | Eq. (3) | Intention-guided variation around a semantic point |
| Cultivate | Interaction history H_t + current z_t | Multi-step, trajectory | Eq. (4) / Algorithm 1 | Bandit-style trajectory optimization driven by preference feedback |
Quick Start
Read and reproduce
- Clone the repository — the official paper (v6, CN/EN, PDF + Word) lives in
paper/. - Open
results/for the latency breakdown, the summary of results (Table 7, the single source of truth) and full hyperparameters (Table A1). - Run the unit-tested reference implementation of Eqs. (2)–(4) and Algorithm 1 in
reference/(numpy only). - Study materials — task protocol, SUS/Likert questionnaires, interview guide, consent and recruitment templates — live in
study/andsupplementary/.
shell
git clone https://github.com/AkiroMusic/Botanica-OS.git cd Botanica-OS # Reproduction environment: Python 3.11 · PyTorch 2.1 · librosa 0.10 · FastAPI # Frontend: React 18 + TypeScript 5 (WebSocket)
研究项目 · 论文与开放材料
项目简介
配套论文的开放材料库。论文提出「有机化操作」设计框架，将神经潜在空间交互重组为三种植物学隐喻范式，在统一的可微 DSP 实时管线上实现原型系统，并通过 N = 12 位 Botanica/IDM 制作人与 ML 研究者的混合方法用户研究量化三范式差异。论文正式版（v6，中英双语，PDF + Word）收录于 paper/ 目录。
论文随库发布
正式版（v6，中英双语，PDF + Word）收录于 paper/ 目录，配 CITATION.cff 引用文件。
三种有机化操作
嫁接——两个音色父本之间的连续过渡；杂交——围绕 CLAP 语义点的意图引导变异；培育——基于偏好反馈的 bandit 式轨迹优化。
混合方法用户研究
N = 12 位制作人与 ML 研究者：嫁接的总体 SUS 最高（78.3），培育的能动性最高（5.9/7），而杂交的实测可控性低于其概念层定位——论文核心发现之一。任务协议、问卷与访谈提纲一并开源。
参考实现
论文式 (2)–(4) 与算法 1 的参考实现，仅依赖 numpy，含单元测试，位于 reference/。
47 毫秒实时管线
CPU 推理端到端延迟 47 ± 3 ms（16 kHz、128 样本块），满足 <100 ms 的演奏式交互要求（DR5）。
完整复现材料
延迟分解、结果总表（表 7）、基线公平性控制、超参数（表 A1）、seeds.yaml 随机种子登记与再分析脚本。
明细
三种有机化操作
| 范式 | 输入 | 时间结构 | 数学形式 | 描述 |
|---|---|---|---|---|
| 嫁接 Graft | 双父本 z_a, z_b + 混合系数 α | 单步、线性 | 式 (2) | 两个音色父本之间的连续过渡 |
| 杂交 Hybridize | 语义条件向量 c（CLAP 文本嵌入） | 单步、概率 | 式 (3) | 在条件语义点周围按意图变异 |
| 培育 Cultivate | 交互历史 H_t + 当前 z_t | 多步、轨迹 | 式 (4) / 算法 1 | 基于偏好反馈的 bandit 式轨迹优化 |
- 克隆仓库——论文正式版（v6，中英，PDF + Word）在
paper/目录。 - 查阅
results/：延迟分解、结果总表（表 7，全文唯一口径）与完整超参数（表 A1）。 - 运行
reference/中式 (2)–(4) 与算法 1 的参考实现（仅依赖 numpy，含单元测试）。 - 研究材料——任务协议、SUS/Likert 问卷、访谈提纲、知情同意书与招募模板——在
study/与supplementary/。
shell
git clone https://github.com/AkiroMusic/Botanica-OS.git cd Botanica-OS # Reproduction environment: Python 3.11 · PyTorch 2.1 · librosa 0.10 · FastAPI # Frontend: React 18 + TypeScript 5 (WebSocket)