FL MiniMax Music 3 Train Config

This node does nothing by itself, and that's the point. It's the settings sheet: every knob you hand it gets packed into a single versioned FL_MINIMAX_MUSIC3_TRAIN_CONFIG object that the LoRA Trainer consumes. No models are loaded, no steps are run - it just translates your choices into a config the pinned SimpleTuner backend can execute. Think of it as the difference between tuning a car and just owning the spec sheet.

The inputs that actually matter

There are a lot of fields, but a beginner genuinely only needs to reason about three of them:

• lora_rank - 16, 32, 64, 128, or 256 (default 64). Higher rank = more capacity to memorize your dataset, but slower and bigger. For a style LoRA from a modest song library, 32–64 is the sane starting band; 256 is for people with big libraries and deep pockets for compute.
• learning_rate - default 5e-05, and honestly the default is a fine place to start. If your LoRA comes out barely audible, the instinct is to crank this - resist it; that's usually a dataset/steps problem, not an LR problem.
• max_train_steps - default 500, but treat that as a smoke-test number. The pack's own example workflow uses deliberately short settings so you can verify the pipeline installs and runs; for real training you'll be pushing this way up.

The rest are worth knowing at a glance rather than obsessing over: lora_alpha (scaling, default 64 - keep it near rank), lora_dropout (default 0), lr_scheduler (cosine/constant/constant_with_warmup/linear/polynomial), lr_warmup_steps, weight_decay, and seed. One pair deserves attention: train_batch_size × gradient_accumulation_steps - the node helpfully reports their product as your effective batch size (default 1 × 4 = 4), which is the number that actually matters for LR sanity.

The VRAM decisions are pre-made for you

The precision fields default to int8-quanto for both base_model_precision and text_encoder_precision. That's the pack trying to keep a ~14 GB model trainable on a single consumer GPU, and for a first run you should leave them alone. no_change is there if you know you have headroom (or are on a big card). gradient_checkpointing defaults on for the same reason. The optimizer options are adamw_bf16 (default), optimi-stableadamw, and bnb-adamw8bit - the bnb option trades a little quality for much lower optimizer VRAM.

checkpoint_interval and checkpoints_total_limit (defaults 100 steps / 3 kept) control how often full checkpoints land and how many are retained - relevant for resume safety, not for output quality.

Install and wiring

This node ships in the FL MiniMax Music 3 pack:

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. In the graph it's a one-liner: Dataset → Train Config → LoRA Trainer. The single train_config output goes into the trainer's required input.

Gotchas

• The config is versioned. The trainer validates that a resumed run matches its original config on dataset identity, LoRA shape, precision, and output name. You can't quietly change rank mid-resume - the pack will refuse, which is a feature, not a bug.
• Don't read 500 steps as "the right number." It's the install-check default. If you actually train a LoRA at 500 steps you'll get a whisper of a style - the common complaint in the music-LoRA community is under-trained LoRAs, not over-trained ones.
• Changing batch size changes effective batch size, and your LR was likely tuned for the default. If you move from 4 to 8 effective batch, scale LR up a touch (or don't - starting from defaults works for most people).