## Technical Guide: Training ACE-Step 1.5 LoRA for Psytrance on an RTX 3090## 1.

Dataset Preprocessing & Surgical Slicing

Standard audio slicing methods destroy the rhythm and phase alignment of dense audio

material like Psytrance. Follow these strict preprocessing rules:

* Zero-Crossing Slicing: Cuts must occur exactly at zero-crossing points ($0\text{ dB}$

amplitude) to avoid digital clicks, pops, and phase cancellation.

* Zero Fade/Crossfade Rule: Never use standard fade-ins or fade-outs. The diffusion model

will interpret this as a musical instruction and learn to fade the track out every 30 seconds.

* 1-Second Crossfade Overlap: Implement a 1-second crossfade overlap between chunks to

maintain continuity across sample boundaries.

* Fixed Chunk Length: Slice the source material into exact 30.0-second segments. This

captures complete musical phrases while fitting comfortably into the 24 GB VRAM limit.

* Format Constraints: Export all slices at 44.1 kHz, 16-bit or 24-bit PCM WAV. Avoid MP3

compression to prevent codec artifacts from muddying the high frequencies.

## 2. Text-Caption Tagging Strategy

Audio diffusion models require metadata to isolate tempo and key. Each audio slice requires

a matching .txt file with identical naming.

* BPM & Key Isolation: Explicitly tag the precise BPM and musical key (e.g., 142 bpm, G#

minor). This prevents the model from blending different tempos and scales into a dissonant

mix.

* Sub-Genre Descriptor: Start every caption with a unified anchor tag (e.g., psytrance track).

* Structural Elements: Document specific sonic elements present in that chunk (e.g., rolling

triplet bassline, punchy energetic kickdrum, sharp acid synth leads, rhythmic percussion,

crisp hi-hats).

* Quality Tokens: Append production quality tags at the end of the text file (e.g., studio

master quality, clean professional mix).

## 3. Training Hyperparameters & VRAM Optimization (RTX 3090)

To maximize the 24 GB VRAM of an RTX 3090 without triggering CUDA out of memory

errors, use these exact network dimensions and pipeline settings:

## Network Architecture (LoRA)

* LoRA Rank ($r$): 64 (Provides sufficient capacity to map distinct keys and tempos into

separate internal slots).

* LoRA Alpha: 32 (Ensures stable weight scaling).

* LoRA Dropout: 0.05 (Prevents overfitting while retaining rapid pattern recognition).

* Target Modules: ["to_q", "to_k", "to_v", "to_out.0", "ff.net.0.proj", "ff.net.2"]

## Optimization & Precision

* Mixed Precision: bf16 (Mandatory for modern GPU compute stability).

* Optimizer: bitsandbytes 8-bit AdamW (Compresses the optimizer states to halve VRAM

allocation).

* Gradient Checkpointing: True (Recomputes activations during the backward pass to save

massive amounts of VRAM).

* Hardware Allocation: Set num_workers=4, pin_memory=True, and

persistent_workers=True.

## Training Schedule

* Batch Configuration: Set train_batch_size: 2 and gradient_accumulation_steps: 2 (Creates

an effective total batch size of 4, ensuring smooth gradient updates for complex audio

signals).

* Learning Rate: 0.00007 ($7\cdot10^{-5}$) with a cosine scheduler and 100 warmup steps.

A lower learning rate preserves sharp transient structures like tight kick drums.

* Seed: Set to -1 (Random Seed) across later epochs to shuffle data blocks and improve

generalization.

## 4. Training Phases & Loss Graph Analysis

The training graph demonstrates a mathematically ideal convergence curve for a dense

audio dataset under a randomized training seed:

Loss

0.55 | \

0.50 | \

0.45 | \_________

0.40 | \________ [Plateau / Saturated Fine-Tuning]

0.35 |______________________

+-----------------------

3100 3300 3500 3700 Step

* Phase 1 (Epoch 0 - 30): Macro-Structure Acquisition: The initial loss drops rapidly from

$\sim0.60$ down to $\sim0.45$. The model identifies coarse structural features, including

noise floors, fundamental frequencies, and the main percussive grid.

* Phase 2 (Epoch 30 - 35): Mid-Frequency Stabilization: The curve forms a gentle slope

between step 3100 and 3400. The random data seed (-1) introduces acoustic variety, forcing

the optimizer to consolidate structural patterns across different BPM/Key signatures

simultaneously.

* Phase 3 (Step 3400 - 3800): Micro-Optimization & Transients: The Loss (smoothed) forms

a textbook plateau between $0.36$ and $0.38$. The raw loss values variance narrows down

significantly, occasionally hitting micro-troughs near $0.31$. This indicates that the model

has fully saturated its learning capacity for the dataset and is purely refining micro-details

like phase alignment and crisp transient sharpness. Pushing the model below $0.30$ is

highly discouraged as it triggers immediate acoustic degradation (overfitting).

## 5. Inference & Audio Generation Configuration

Once training concludes at Epoch 40, halt the script and configure the Inference tab using

these precise generation parameters:

* Inference Backend: Set to PyTorch (Do not use vLLM or Triton on native Windows

environments due to library compatibility issues).

* Base Model Path: Point to checkpoints/acestep-v15-xl-sft.

* LoRA Model Path: Load the target checkpoint (e.g., epoch_35 or epoch_40).

* LoRA Scale: 0.85 to 1.0 (Start at 0.85 to maintain flexibility; increase to 1.0 if the synthetic

output lacks the driving weight of the original data).

* Inference Steps: 50 (Provides clean diffusion generation without blurring the fast

transients).

* CFG Scale: 4.5 to 5.5 (Higher values force strict adherence to the prompt tags, lower

values add acoustic variation).

* Audio Length: Exact 30.0 seconds (Must match the training slice length; generating beyond

this window causes structural collapse).

* Target Generation Prompt: Feed the explicit tokens used during tagging to extract the

clean, isolated style:

A high-energy psychedelic trance track, 142 BPM, fast driving rolling bassline, punchy

energetic kickdrum, sharp acid synth leads, rhythmic percussion, crisp hi-hats, studio master

quality, clean professional mix