## Technical Guide: Training ACE-Step 1.5 LoRA for Psytrance on an RTX 3090## 1.Dataset Preprocessing &amp; Surgical SlicingStandard audio slicing methods destroy the rhythm and phase alignment of dense audiomaterial 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 modelwill 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 tomaintain continuity across sample boundaries.* Fixed Chunk Length: Slice the source material into exact 30.0-second segments. Thiscaptures 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 MP3compression to prevent codec artifacts from muddying the high frequencies.## 2. Text-Caption Tagging StrategyAudio diffusion models require metadata to isolate tempo and key. Each audio slice requiresa matching .txt file with identical naming.* BPM &amp; 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 dissonantmix.* 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., rollingtriplet 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., studiomaster quality, clean professional mix).## 3. Training Hyperparameters &amp; VRAM Optimization (RTX 3090)To maximize the 24 GB VRAM of an RTX 3090 without triggering CUDA out of memoryerrors, use these exact network dimensions and pipeline settings:## Network Architecture (LoRA)* LoRA Rank ($r$): 64 (Provides sufficient capacity to map distinct keys and tempos intoseparate 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", "<a target="_blank" rel="ugc" href="http://ff.net">ff.net</a>.0.proj", "<a target="_blank" rel="ugc" href="http://ff.net">ff.net</a>.2"]## Optimization &amp; Precision* Mixed Precision: bf16 (Mandatory for modern GPU compute stability).* Optimizer: bitsandbytes 8-bit AdamW (Compresses the optimizer states to halve VRAMallocation).* Gradient Checkpointing: True (Recomputes activations during the backward pass to savemassive amounts of VRAM).* Hardware Allocation: Set num_workers=4, pin_memory=True, andpersistent_workers=True.## Training Schedule* Batch Configuration: Set train_batch_size: 2 and gradient_accumulation_steps: 2 (Createsan effective total batch size of 4, ensuring smooth gradient updates for complex audiosignals).* 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 improvegeneralization.## 4. Training Phases &amp; Loss Graph AnalysisThe training graph demonstrates a mathematically ideal convergence curve for a denseaudio dataset under a randomized training seed:Loss0.55 | \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, includingnoise floors, fundamental frequencies, and the main percussive grid.* Phase 2 (Epoch 30 - 35): Mid-Frequency Stabilization: The curve forms a gentle slopebetween step 3100 and 3400. The random data seed (-1) introduces acoustic variety, forcingthe optimizer to consolidate structural patterns across different BPM/Key signaturessimultaneously.* Phase 3 (Step 3400 - 3800): Micro-Optimization &amp; Transients: The Loss (smoothed) formsa textbook plateau between $0.36$ and $0.38$. The raw loss values variance narrows downsignificantly, occasionally hitting micro-troughs near $0.31$. This indicates that the modelhas fully saturated its learning capacity for the dataset and is purely refining micro-detailslike phase alignment and crisp transient sharpness. Pushing the model below $0.30$ ishighly discouraged as it triggers immediate acoustic degradation (overfitting).## 5. Inference &amp; Audio Generation ConfigurationOnce training concludes at Epoch 40, halt the script and configure the Inference tab usingthese precise generation parameters:* Inference Backend: Set to PyTorch (Do not use vLLM or Triton on native Windowsenvironments 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 syntheticoutput lacks the driving weight of the original data).* Inference Steps: 50 (Provides clean diffusion generation without blurring the fasttransients).* CFG Scale: 4.5 to 5.5 (Higher values force strict adherence to the prompt tags, lowervalues add acoustic variation).* Audio Length: Exact 30.0 seconds (Must match the training slice length; generating beyondthis window causes structural collapse).* Target Generation Prompt: Feed the explicit tokens used during tagging to extract theclean, isolated style:A high-energy psychedelic trance track, 142 BPM, fast driving rolling bassline, punchyenergetic kickdrum, sharp acid synth leads, rhythmic percussion, crisp hi-hats, studio masterquality, clean professional mix

ACE-Step 1.5 Psytrance LoRA Ver 2.0

pretrained_model_path: "checkpoints/acestep-v15-xl-sft"output_dir: "output/psytrance_lora_xl_sft"# --- DATASET ---dataset:path: "PSY_MASTER_DATASET"sample_rate: 44100slice_duration: 30.0# --- HYPERPARAMETER ---train_batch_size: 2gradient_accumulation_steps: 1epochs: 30learning_rate: 0.00007lr_scheduler: "cosine"lr_warmup_steps: 100# --- LORA PARAMS ---lora_rank: 64lora_alpha: 32target_modules: ["to_q", "to_k", "to_v", "to_out.0", "<a target="_blank" rel="ugc" href="http://ff.net">ff.net</a>.0.proj", "<a target="_blank" rel="ugc" href="http://ff.net">ff.net</a>.2"]# --- OPTIMIERUNGEN ---mixed_precision: "bf16"gradient_checkpointing: trueoptimizer: "adamw_8bit"# --- SAVING ---save_every_epochs: 5checkpointing_steps: 4000This is a high-quality Psytrance Audio LoRA trained on ACE-Step 1.5 SFT-XL. It is optimized for generating punchy rolling triplet basslines, crisp percussion, and sharp acid synth leads.Training Details:- Base Model: ACE-Step 1.5 XL SFT- Dataset: 44.1 kHz Studio Masters (30-second seamless zero-crossing slices)- Hyperparameters: Rank 64, Alpha 32, Learning Rate 7e-5Recommended Inference Settings:- Inference Steps: 50- CFG Scale: 4.5 - 5.5- LoRA Scale: 0.85 - 1.0- Audio Length: Exactly 30.0 seconds per chunk___________________________________________________________"caption": "psytrance style, , atmospheric pads,rolling bassline, acid squelch, metallic leads, high quality, ", "global_caption": "", "lyrics": "[Intro - Atmospheric Pads &amp; Ambient Synth]\n\n", "instrumental": false, "vocal_language": "unknown", "bpm": 135, "keyscale": "", "timesignature": "4", "duration": 30, "enable_normalization": true, "normalization_db": -1, "fade_in_duration": 0.0, "fade_out_duration": 0.0, "latent_shift": 0, "latent_rescale": 1, "inference_steps": 120, "seed": 1297183202, "guidance_scale": 8.1, "use_adg": false, "cfg_interval_start": 0, "cfg_interval_end": 1, "shift": 3, "infer_method": "ode", "sampler_mode": "heun", "velocity_norm_threshold": 0, "velocity_ema_factor": 0, "dcw_enabled": true, "dcw_mode": "double", "dcw_scaler": 0.02, "dcw_high_scaler": 0.06, "dcw_wavelet": "haar", "timesteps": null, "repainting_start": 0, "repainting_end": -1, "chunk_mask_mode": "auto", "repaint_latent_crossfade_frames": 10, "repaint_wav_crossfade_sec": 0.0, "repaint_mode": "balanced", "repaint_strength": 0.5, "retake_seed": null, "retake_variance": 0.0, "flow_edit_morph": false, "flow_edit_source_caption": "", "flow_edit_source_lyrics": "", "flow_edit_n_min": 0.0, "flow_edit_n_max": 1.0, "flow_edit_n_avg": 1, "audio_cover_strength": 0.93, "cover_noise_strength": 0, "thinking": true, "lm_temperature": 0.85, "lm_cfg_scale": 2, "lm_top_k": 0, "lm_top_p": 0.9, "lm_negative_prompt": "bitcrushed, aliasing, quantizing noise, digital clipping, glitchy, stutter, stuttering, dropouts, artifacting, mp3 artifacts, 64kbps, encoded,jazz, funk, pop, acoustic, lo-fi, orchestral,house, techno, dubstep, pop, vocal hooks, acoustic instruments, electric guitar, slow tempo, jazz chords, ambient drone, lo-fi hiss, distorted drums, orchestral elements, trap hi-hats,major scale, happy chords, uplifting melody, bright pop progression, standard minor chord changes, blues scale", "use_cot_metas": false, "use_cot_caption": false, "use_cot_lyrics": false, "use_cot_language": false, "use_constrained_decoding": true, "cot_bpm": null, "cot_keyscale": "", "cot_timesignature": "", "cot_duration": null, "cot_vocal_language": "unknown", "cot_caption": "", "cot_lyrics": "", "lora_loaded": true, "use_lora": true, "lora_scale": 1.0, "lora_weights_hash": "0c942944d792d52643a33c2e54b236c2ad6a4e26af46503870abbedb26178490", "audio_format": "mp3", "mp3_bitrate": "320k", "mp3_sample_rate": 44100, "repaint_source_latents_file": "e5c0775d-18f4-0022-de7f-a5635c521ed2.repaint_latents.npy", "session_artifact_file": "e5c0775d-18f4-0022-de7f-a5635c521ed2.session.npz", "session_artifact_kind": "generation_intermediates_v1"}

Files with SHA256: febabf7a9d388591290ae08710b9bd351c79f2f436f77c14df4caf08845cb8af