FL MiniMax Music 3 Validation Config

Watching a loss curve drop tells you training is happening, but it tells you nothing about whether the music sounds like your dataset. This node gives you the thing a loss chart can't: actual audio. It defines a deterministic validation track - same prompt, same seed, same settings every time - that the trainer generates at fixed intervals while your LoRA trains, so you can listen to the same 15 seconds of music slowly turn into your style. It's the difference between guessing and knowing, and for music LoRAs specifically, it's the closest thing to a north star.

Why deterministic matters

The whole trick is that the validation generation is reproducible. Same caption, same lyrics, same seed, same inference steps, same guidance - the only thing changing between generations is the LoRA itself. That's what makes the before/after comparison honest. If the seed varied, you'd be comparing apples to oranges and learning nothing. The pack writes these samples out with each checkpoint interval, so you can scrub through them like a film strip of your LoRA's progress.

The inputs

• caption - the musical description, defaulting to something like funky house groove, 125 bpm, warm bassline, crisp drums. Change it to match the style you're training, or leave it as a neutral canary.
• lyrics - optional; leave empty for instrumental validation, or write a short vocal line to check how the LoRA handles the lyric side of Music 3's dual conditioning.
• duration - 5 to 60 seconds (default 15). Longer is more representative but costs validation time; 15 is a good compromise.
• seed - fixed default 42. Keep it fixed - that's the point.
• inference_steps - default 30, the sampler's step count.
• guidance - default 1.7. That low number is normal for Music 3's diffusion setup, not a typo.
• validation_interval - how many training steps between validations (default 100). Ties neatly to the checkpoint interval.
• samples - 1 to 8. More samples = a fuller listen (or more seeds), at the cost of more validation compute.

The single validation_config output feeds the LoRA Trainer's optional validation_config input. Nothing else consumes it; it's purely a mid-training instrument.

Install

Part of the FL MiniMax Music 3 pack - ComfyUI Manager (search FL MiniMax Music 3) or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxMusic3.git
cd ComfyUI-FL-MiniMaxMusic3
pip install -r requirements.txt

Restart ComfyUI, then wire Dataset → Train Config → LoRA Trainer with this node's output hanging off the trainer's optional input.

Honest expectations

Validation runs eat GPU time mid-training - every validation_interval steps, the trainer pauses to generate and decode audio. With the default 15-second sample it's modest, but set samples to 8 at 60 seconds and you'll notice. Also calibrate what you're listening for: early checkpoints will sound like the base model with a faint accent, and that's correct. If you're chasing an audible style, resist the urge to validate every 50 steps and instead fix the dataset. Also: it validates what the LoRA sounds like at this checkpoint, not what a finished multi-minute track sounds like - treat it as a progress meter, not the final review.