MiniMax Music 3 Latent Refiner¶
Development line: project:minimax-music3-latent-refiner · thread minimax-music3-latent-refiner
Last event: 2026-08-23 · 1 dated since 2026-08-23 · Researched: 2026-09-05 · confidence: medium
What it is¶
MiniMax Music 3 Latent Refiner is a community checkpoint for restoring existing audio, not a text-to-music generator.
- Audio repair: cleans damaged audio tracks.
- Fidelity: preserves performance, timing, vocals, and arrangement.
- Runtime: runs through MiniMax Music 3 DAV.
- Interfaces: available via CLI, Python, Diffusers attachment, and ComfyUI.
Tested only on CUDA with FP32; severe damage outside the training pipeline destroys unrecoverable information.
Use it as post-processing for MiniMax Music 3 audio with listening checks, never as a replacement for original recordings or a text-to-music pipeline.
Development line¶
- 2026-08-23 — MiniMax Music 3 Latent Refiner v0.10 was linked on Hugging Face. On 2026-08-23, a Hugging Face link for BornSaint/minimax-music3-latent-refiner-v0.10 was recorded. The versioned artifact link indicates a concrete v0.10 step in the MiniMax Music 3 Latent Refiner development line, although its capabilities and changes were not established from the available evidence.
What changed¶
- 2026-08-23 — The page
BornSaint/minimax-music3-latent-refiner-v0.10appeared, but the saved link alone does not confirm a working release or its specifications. - 2026-09-01 — Published v0.10 at
terminusresearch/minimax-music3-latent-refiner-v0.10: a 137,253,888-parameter bridge-hybrid checkpoint-1000 for restoring audio in DAV latents.
How to use this¶
As of 2026-08-23, treat the linked v0.10 artifact as a candidate version to inspect and evaluate, rather than assuming an earlier latent-refiner version remains current.
- Clone the repository with Git LFS, create a Python environment, and install the package in editable mode. — https://github.com/bghira/minimax-music3-latent-refiner
- Run
minimax-music3-refine input.wav refined.wavfor basic file processing. — https://github.com/bghira/minimax-music3-latent-refiner - For Python, load
MiniMaxMusic3RefinerPipeline, pass waveform and sample rate, then save the returned audio tensor. — https://github.com/bghira/minimax-music3-latent-refiner - For ComfyUI, install the custom node, restart ComfyUI, and open the included workflow; use the refiner loader, DAV encoder loader, and latent refine node. — https://github.com/bghira/minimax-music3-latent-refiner
Best practices¶
- Use overlapping 30-second windows with a two-second latent overlap: the checkpoint was trained on 30-second windows, which is the stated quality baseline. — https://github.com/bghira/minimax-music3-latent-refiner
- Prefer CUDA and FP32: this is the verified release precision; do not treat the bfloat16 example in the auto-generated Hub instructions as a confirmed mode for this audio pipeline. — https://github.com/bghira/minimax-music3-latent-refiner
- Check the result by ear, and do not expect recovery of information destroyed by severe damage outside the training pipeline. — https://github.com/bghira/minimax-music3-latent-refiner
Superseded by this¶
- 2026-08-23 — The instructions on the
BornSaint/minimax-music3-latent-refiner-v0.10page showing an image prompt andpipe(prompt).images[0]are not a valid way to run an audio refiner. - 2026-09-01 — The active release identifier in the source code and model card is
terminusresearch/minimax-music3-latent-refiner-v0.10; do not substitute the BornSaint page of the same name.
Still unknown¶
BornSaint/minimax-music3-latent-refiner-v0.10currently shows an inconsistent image-generation snippet, though it links toterminusresearchin the ComfyUI section. Without a historical snapshot, we cannot verify what was published there on 2026-08-23.- The v0.10 release is confirmed by a commit shown four days before the 2026-09-05 check; the 2026-09-01 date is derived from this relative timestamp.
- No independent quality benchmarks, checks across damage types, or VRAM measurements exist.
Sources¶
| source | title | read |
|---|---|---|
| https://huggingface.co/BornSaint/minimax-music3-latent-refiner-v0.10 | BornSaint/minimax-music3-latent-refiner-v0.10 — Hugging Face | 2026-09-05 |
| https://github.com/bghira/minimax-music3-latent-refiner | bghira/minimax-music3-latent-refiner — GitHub | 2026-09-05 |
| https://huggingface.co/terminusresearch/minimax-music3-latent-refiner-v0.10/commit/8806b37005e090f1b3443759aa03563a5105d4d9 | Release MiniMax Music 3 latent refiner v0.10 — Hugging Face commit 8806b37 | 2026-09-05 |
Agent brief¶
- Subject:
project:minimax-music3-latent-refiner, threadminimax-music3-latent-refiner, 1 dated events 2026-08-23 → 2026-08-23. - Practical note: As of 2026-08-23, practitioners should treat the linked v0.10 artifact as a candidate version to inspect and evaluate, rather than assuming an earlier latent-refiner version remains current.
- Confidence: medium. Dated supersedes above are the authority for what is obsolete.