Encode the inputs
The VAE turns backing audio into continuous latents. A separate text encoder represents the instruction and caption.
Hear the same 30 seconds.
e06 · Base · 114514A fixed bass-and-drums backing, a generated guitar candidate, then the two together.
Backing only00:00 — 00:30
Switching keeps your position. Each isolated part has the same gain as in the mix; no separate loudness normalisation.
Play the Base 30-second waveform demonstration
An offline visualisation of the saved Base result, with its mixed audio. It is not a recording of live generation.
Why this is better. Why the other falls short.
“Better” and “worse” refer to these two examples and the requested guitar role. Neither is a finished musical solution.
| Question | Base · e06 | XL-Base · e04 OFF |
|---|---|---|
| Does it sound like the requested instrument? | The guitarist-researcher recognises distorted electric-guitar tone and plausible playing character in the mix. | The researcher says the previous problems remain and prefers Base in this pair; no new detailed XL score was supplied. |
| What does the less convincing example still achieve? | A positive lead-guitar character judgement; phrasing and harmonic fit are still unassessed. | Earlier solo listening identified an audible, single-note melody. A single-note line is not itself a guitar-solo failure. |
| Does the part enter when needed? | 0–10 s raw RMS: −51.70 dBFS. A very quiet opening remains. | 0–10 s raw RMS: −57.95 dBFS. The same low-energy opening remains. |
| What can the spectrum support? | 93.67% of analysed band energy lies in 250–2000 Hz: less concentrated than XL in this pair. | 98.05% lies in 250–2000 Hz. This is an acoustic difference, not proof of instrument identity or musical quality. |
| Is the file technically intact? | 30 s, stereo, 48 kHz; finite samples and no full-scale samples. | 30 s, stereo, 48 kHz; finite samples and no full-scale samples. |
| What remains unverified or unresolved? | A convincing isolated part, explicit rhythmic/harmonic fit and reliable results across inputs are not yet established. The cause of the shared delayed entrance is unresolved. | |
RMS and spectral values describe the raw model outputs before audition gain. Band fractions use 20–12,000 Hz as the denominator after 24 kHz analysis resampling. These measurements support acoustic observations; they do not replace listening.
View the existing waveform and spectrum evidence
The normalised onset plot shows relative feature changes, not absolute loudness or a musical-alignment score.
Both examples use direct LEGO with DCW OFF, the same backing, prompt, seed and explicit generation settings. The checkpoint changes; the XL reference was generated earlier. Equal seeds do not establish equal internal noise or a causal explanation of the quality difference.
The Base request, with generation settings matched to XL.
Direct source audio · no music-LM codesINPUT 1 / AUDIO
30 seconds of bass + drums
Extraordinary, original 64–94 s. Guitar stems were excluded when constructing the backing. The full backing file entered the generation request as src_audio.
- src_audio
- backing.wav
- reference_audio
- null
- audio_codes
- "" (empty)
- BPM / key / metre
- not supplied
Tempo and key estimates from a separate understanding call were not used. Musical context came from the audio.
The generation used the full-level source file. Audition copies use 0.5 gain to match each part’s level in the fixed mix.
INPUT 2 / EXACT TEXT SENT
Generate the GUITAR track based on the audio context:
Solo lead electric guitar, foreground melodic phrases with pauses, not rhythm-guitar strumming. Guitar only, without drums, bass, vocals or other instruments.
These English strings are the actual request, unchanged by the language switch. There was no recovery, moderate-energy or game-scene prompt.
Generate a new part
LEGO supplies source context and the target-track instruction to the Base diffusion transformer (DiT), which generates the candidate representation.
Decode the candidate
The VAE decodes the generated representation to a 30-second waveform. “Guitar only” is the requested target; stem purity still needs listening.
Mix for listening
Outside the model: 0.5 × backing + 0.5 × candidate. The supplied backing is retained. This mix is not a second generation.
acestep-v15-baseLEGO50 steps · Euler / ODEseed 114514CFG 7 · shift 3DCW OFF
Inspect the request and execution evidence
Verified against the saved run
- The loaded checkpoint was acestep-v15-base, with VAE and Qwen3-Embedding text encoder.
llm_initialized=False · use_lm=FalseThe music LM was not loaded. The text encoder still ran: text encoding is not music-code generation.
- The log records source-audio processing. Empty audio_codes allows this direct source path.
- One generation call, no inference retry. Compared with the reused XL OFF sample, explicit generation settings were matched and the checkpoint changed.
Selected request fields plus a full portable parameter record. Source paths use portable file names; original records remain in the project.
Use the same questions for the next example.
Current case: Base / XL direct LEGODefine success before judging
Name the input, target part and musical purpose. Judge instrument character, timing, context fit and unwanted content separately.
Trace what actually ran
Preserve the exact instruction, audio and parameters. Check which conditions reached the model and which requested operations executed.
Show strengths and failures
Compare the same passage as backing, isolated part and mix. Include convincing and weaker examples, with their successful and unsuccessful aspects.
Tie each conclusion to evidence
Use listening for musical judgements and signal analysis for acoustic observations. Keep unknowns distinct from failures; test a proposed cause before treating it as the explanation.
How this case exposed the need for the method
- 01e01 / e02
Make the question audible
Generate a foreground part over an existing backing, on the same timeline. Fix the bass-and-drums input and retain three routes.
- 02Input audit
Check what each route consumes
Direct audio, source-encoded codes and text-LM codes do not provide equivalent context. Non-empty codes take precedence over source audio; the text-LM route did not hear the backing. Remove conflicting mood wording and describe only the target part.
- 03e04 · OFF / ON
Repair an operation that had been skipped
DCW is a wavelet correction during latent sampling. It was requested ON but skipped after a dependency import failure. After repair, effective ON changed the output but the listened example remained noisy. The execution issue was fixed; musical quality was not solved.
- 04P0 · e05
Use the supported task; question the analysis
Both Base and XL-Base support LEGO. A separate audio-understanding call estimated 40 BPM, A minor and a guitar riff. Those unverified estimates were not fed into generation.
- 05P1 · e06
Try Base with the direct audio route
Keep the explicit generation settings from the XL OFF reference and change the checkpoint. The new Base example receives positive mixed feedback on guitar timbre and playing character. This is a local result, not a universal model ranking.