ace-step: gallery entry for acestep-v15-xl-base + expose audio-conditioned task types (cover/repaint/lego) in /v1/sound-generation and Studio Sound UI #12721

Is your feature request related to a problem?

The ace-step backend supports only text-to-music via the /v1/sound-generation endpoint and the Studio "Sound" page. The upstream ACE-Step 1.5 model family — and the Python backend implementation in
backend/python/ace-step/backend.py — already implements the full task matrix (text2music, cover, repaint, extract, lego, complete) and accepts audio-conditioned inputs, but none of it is reachable from the HTTP API or
the React UI.

Describe the solution

Four related improvements:

1. Gallery entry for acestep-v15-xl-base (4B DiT)

ace-step-turbo is already in the model gallery; please add the XL base variant (ACE-Step/acestep-v15-xl-base), the highest-quality DiT of the family (~18.8 GB bf16, MIT license). Manual setup works today but requires hand-placing weights under models/<name>/checkpoints/<name>/.

2. LM variant selection in the gallery entry

The backend supports lm_model_path with acestep-5Hz-lm-0.6B, acestep-5Hz-lm-1.7B and acestep-5Hz-lm-4B. The gallery entry (or two variants of it) should let the user pick which LM is downloaded and
configured, since the 4B planner meaningfully improves CoT metadata/caption quality.

3. Audio Reference inputs on /v1/sound-generation + Studio Sound UI

The backend already maps request payload keys into GenerationParams, including task_type, reference_audio_path, src_audio, audio_cover_strength and repainting_start/repainting_end. What is
missing is the plumbing above it:

• a way to upload/pass a source audio file in /v1/sound-generation (e.g. multipart form or a file-path field);
• UI fields on the Studio Sound page for:
  • reference_audio — global timbre/style reference (cover, style transfer);
  • src_audio — source audio for non-text2music tasks;
  • audio_codes — semantic codes for cover-mode reuse/extensions;
  • repainting_start / repainting_end — interval control for repaint and lego tasks.

4. task_type selector with task-specific controls

Expose task_type in the API and add a selector in Studio Sound with the ACE-Step 1.5 task types: text2music, cover, repaint, lego, extract, complete — enabling the relevant inputs per task (lyrics and audio refs for cover, intervals for repaint/lego, etc.).

Additional context

• Backend capability already exists: backend/python/ace-step/backend.py forwards model options into the generation payload, and GenerationParams accepts the fields above — this feature is endpoint + UI plumbing, not new inference work.
• The audio.cpp backend already has a source_audio input path for editing routes (build_sound_generation_request), so this would align the Python ace-step backend with a pattern the codebase partially supports.

Environment

• LocalAI v4.11.0 (binary), cpu-ace-step backend
• ACE-Step 1.5 XL base + acestep-5Hz-lm-4B working via manual YAML