FL AceStep Training Configuration

This node doesn't do any work. It packages hyperparameters into a config object and hands it to FL AceStep Train LoRA. But it's where runs get won or lost, and the defaults are honestly pretty sane for a first attempt. Expect to spend your time here - it's the one node in the pack that rewards actually reading the tooltips.

The inputs that matter

• lora_rank (default 8, range 4–256, step 4) - capacity of the adapter. 8 is small and safe. The image-LoRA instinct says go bigger, but audio style LoRAs work fine at low rank, and it's far easier to overtrain music than undertrain it. Real-world ACE-Step training mostly lives in the 8–32 band.
• lora_alpha (default 16, range 4–512) - the scaling factor. The default of 2× rank is a good ratio to keep as you move rank.
• lora_dropout (default 0.1, range 0–0.5) - regularization. The 0.05–0.2 band is typical.
• learning_rate (default 1e-4, range 1e-6–1e-2) - 1e-4 is the right starting point for bf16 on this model. If your loss explodes, drop it a notch; don't chase the ceiling.
• max_epochs (default 100, range 10–10000) - the overtraining dial, and the source of the pack's most common complaint. The community's loudest failure mode with this pack is blasting 1000–1600 epochs on a handful of songs and ending up with a LoRA you can barely hear at strength 1.0 (there's a whole "failing spectacularly" thread about exactly that - it wasn't the model, it was the epochs). Start at 50–100, watch the loss, and let checkpoints do the rest.
• batch_size (default 1, range 1–8) - leave it at 1 on a consumer card.
• gradient_accumulation (default 4, range 1–16) - your effective-batch dial. Effective batch = batch_size × accumulation. Raise this before you raise batch_size.
• save_every_n_epochs (default 10) - checkpoint cadence. Lower it while you're tuning so you get a ladder of checkpoints to pick from instead of one final.
• output_dir (default ./output/acestep/loras) - where trained LoRAs land.
• seed (default 42) - set it, keep it reproducible.

The optional row - warmup_steps (100), weight_decay (0.01), max_grad_norm (1.0) - you can leave alone until you're debugging. target_modules (q_proj,k_proj,v_proj,o_proj) is which attention projections get LoRA injected; the default covers the standard four projections and you'll rarely need to touch it.

What's fixed, and why it's good

Mixed precision is bf16 - there's no fp32 option. That's partly why this pack's LoRAs come out roughly half the size of the official ACE-Step UI's fp32 adapters at the same rank, and why people report faster training with it. The turbo model trains on the fixed 8-step discrete timesteps with shift 3.0 - the pack handles that internally, so you don't think about timesteps at all.

Output

A single config output (ACESTEP_TRAINING_CONFIG), wired into the Train node. That's the whole contract.

Installing it

Pack-level install, same as every node in this set:

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

Restart ComfyUI, or use ComfyUI Manager and search "FL AceStep".

The one piece of advice worth stealing

The LoRA-training crowd has a durable rule that maps straight onto audio: fewer, cleaner samples and reasonable epochs beat a huge dataset run into the ground. ACE-Step style training works from "a few songs" - the pack's own README and community reports both land there. So: low rank, 1e-4, batch 1 with accumulation, 50–100 epochs, save every 10, and iterate on the checkpoints rather than cranking the epoch count to 5000. Your ears will thank you.