Generate fully expanded · mobile · error states

Last batch. Generate tab with every control surfaced. Mobile phone screens for Generate + Cover + Lyrics. Six error/edge-case states.

🎵 Generate — fully expanded · psytrance preset stacked with custom LoRA

ACE Music Studio.
Built with . Drop a like · Follow @techfreakworm for what's next.
ready · MPS · M5 Max
🎵Generate
🎤Cover
Extend
✏️Edit
✍️Lyrics
History · session
▶ psytrance · just now
▶ ambient_v4 · 2m
▶ chinese_rap · 7m
▶ lofi_vocal · 14m
1 · Style prompt describe the song · genre, instruments, mood
psytrance, rolling triplet bassline, acid squelch, metallic leads, atmospheric pads, high quality
2 · Lyrics use [verse] [chorus] [bridge] tags · ↗ open Lyrics tab to draft with Qwen 2.5
[intro - atmospheric pads & ambient synth]

[verse 1] six in the morning, the sun's still pretending
kick drum carries what the night was sending
shoes off, eyes closed, the city's still bending

[chorus] we let go, we let go, we let go
Duration 5 – 240 s
seconds30
Vocal mode
With vocals
Instrumental
LoRA stack · 2 active · order matters
Bundled presets click to toggle
RapMachine Chinese Rap Lyric2Vocal Text2Samples
Active stack ↑↓ to reorder · × to remove
Lyric2Vocal · preset · 28 MB 0.65 ×
psytrance_v2 · custom · 47 MB · rank 64 · sha 0c94… 0.95 ×
↑ drop .safetensors here or click
Advanced · generation parameters
BPM
135
Key / scale
auto
Time signature
4 / 4
Sampler
heun
Vocal language
auto
Inference steps50
CFG scale5.0
Shift3
CFG interval start0.0
CFG interval end1.0
Negative prompt things to avoid
bitcrushed, aliasing, quantizing noise, digital clipping, glitchy, mp3 artifacts, jazz, funk, pop, acoustic, lo-fi, orchestral, dubstep, vocal hooks, electric guitar, slow tempo, jazz chords, blues scale
Audio format
mp3 320
wav 44.1
Loudness
-14 LUFS
Fade in
seconds0.0
Fade out
seconds0.0
Latent shift
0
Latent rescale
1
Seed
1297183202
 
Lock seed
LM planner · Qwen3 thinking · chain-of-thought structure
Thinking enabled + slower but better structure
Constrained decoding
Temperature
0.85
Top-k
0
Top-p
0.90
LM CFG
2
CoT pipeline toggles which fields the LM rewrites pre-generation
metas
caption
lyrics
language
LM negative prompt
happy chords, major scale, uplifting melody
CoT override fields if a CoT toggle is on, the LM rewrites these
cot_bpm
(blank → use main BPM)
cot_keyscale
(blank → use main key)
DCW · dynamic CFG warping · wavelet-based
DCW enabled
Mode
double
Wavelet
haar
 
leave defaults if unsure
DCW scaler0.02
High scaler0.06
▶ Generate · est. ~30 s on M5 Max
Output · psytrance · 30 s · seed 1297183202
0:00 / 0:30 ↻ retake · new seed
Stems · Demucs htdemucs_ft
vocals · 1.8 MB
drums · 1.6 MB
bass · 1.4 MB
other · 1.7 MB
Export
↓ mp3 · 1.2 MB ↓ wav · 5.3 MB ↓ stems zip { } meta ↗ share
Metadata
{
  "mode": "generate",
  "prompt": "psytrance, rolling triplet bassline...",
  "lyrics_first_line": "[intro - atmospheric pads...",
  "duration_s": 30, "instrumental": false,
  "bpm": 135, "key": "auto", "time_sig": "4/4",
  "sampler": "heun", "steps": 50, "cfg": 5.0, "shift": 3,
  "cfg_interval": [0.0, 1.0],
  "lm": {"thinking": true, "temp": 0.85, "top_p": 0.9, "cfg": 2,
       "cot": {"metas":false,"caption":false,"lyrics":false,"language":false}},
  "dcw": {"enabled":true,"mode":"double","scaler":0.02,"high_scaler":0.06,"wavelet":"haar"},
  "loras": [
    {"name":"Lyric2Vocal","scale":0.65,"sha256":"7e1f..."},
    {"name":"psytrance_v2","scale":0.95,"sha256":"0c94..."}
  ],
  "seed": 1297183202,
  "output_sha256": "f33a..."
}

📱 Mobile — phone screens

Horizontal scroll tab strip at the top replaces the sidebar. Output stacks below form. Same Brutalist Mono.

ACE Music.
♥ @tfw
🎵 Generate
🎤 Cover
✏️
✍️
Style
psytrance, acid leads
Lyrics
[verse] six in the morning...
Duration · 30 s
psytrance_v2 + upload
▸ Advanced · BPM 135, sampler heun
▸ LM planner
▸ DCW
▶ Generate
0:00 / 0:30
↓ mp3 ↓ wav stems
Generate · 360 × 720 mobile
ACE Music.
♥ @tfw
🎵
🎤 Cover
✏️
✍️
1 · Reference
↑ ref_psy.wav
44.1k · 28 s · 2.1 MB
2 · New prompt
faster, more aggressive
3 · New lyrics
[verse] new lyrics over ref...
Cover strength · 0.93
RapMachine
▸ Advanced
▸ LM planner
▶ Cover
Cover · with ref audio loaded
ACE Music.
♥ @tfw
🎵
🎤
✏️
✍️ Lyrics
Brief
psytrance anthem about sunrise...
Structure
intro, verse, chorus...
Language · en · 0.85 temp
▸ LM parameters
▶ Draft
[intro]
the lights start low...
[verse 1]
six in the morning,
the sun's still pretending...
↑ Use in Generate
Lyrics · draft visible

⚠️ Error and edge-case states

!
LoRA not compatible LoRAValidationError
This LoRA was trained against SDXL, not ACE-Step 1.5 XL SFT. Expected DiT modules: to_q, to_k, to_v, to_out.0, ff.net.0.proj, ff.net.2. Got: unet.down_blocks….
Remove from stack
View header diagnostics
ZeroGPU timed out · auto-retried at 2× duration
First attempt aborted at the 60 s shared-A10G cap. Second attempt at 120 s also aborted. Try a shorter duration, fewer steps, or fewer active LoRAs. last seen: 70 s wall, step 41/50
Lower steps to 30
Reduce duration to 20 s
MPS op not implemented · falling back to CPU aten::_fft_r2c
An ACE-Step kernel hit a PyTorch MPS gap. CPU fallback engaged via PYTORCH_ENABLE_MPS_FALLBACK=1. Generation will continue but be ~2–3× slower for the affected segments.
Continue anyway
Open issue on GitHub
!
Reference audio rejected unsupported format
Cover mode needs wav, mp3, or flac, ≤ 60 s, ≤ 50 MB. Got m4a, 4:12 long, 87 MB.
Pick a different file
Auto-convert + trim to first 60 s
i
First request — warming up the pipeline (~45 s)
Loading ACE-Step v1.5 XL SFT weights into MPS memory. Subsequent generations in this session start instantly.
Generating… step 21 / 50 · ETA 14 s
heun sampler · CFG 5.0 · 2 LoRAs active · seed 1297183202
0:08 elapsed ↻ cancel

All mockups approved — bake them into the spec

Move every approved mockup into docs/superpowers/specs/mockups/ and reference them from §8 of the spec. Then hand off to writing-plans.

Revise something specific

Tell me which mockup / control / error needs work.