AceStep 1.5 SFT Generate Advanced

If you've ever split a 50-step image generation across two KSampler Advanced nodes, you already understand this node. It's Generate with the noise and step-range controls exposed: add_noise, noise_seed, start_at_step, end_at_step, return_with_leftover_noise. Everything else - the sampler list including the bundled jkass_quality, the guidance controls, the outputs - behaves exactly as the regular Generate does.

It exists for two-stage sampling - run half the steps, inspect the latent, finish with different settings - and for pipelines that hand a latent from one stage to the next. What it is not is a "better" Generate: the author says it plainly, splitting stages "adds control, not a guaranteed quality gain."

What changes versus Generate

Under the hood it's a subclass that strips denoise (there's no single denoise value when you're slicing a schedule) and removes the task input entirely, so native Cover and Inpaint tasks must use regular Generate. It then calls the same sampler with a full denoise of 1.0, disable_noise from your add_noise choice, and force_full_denoise from return_with_leftover_noise. Your steps value still defines the whole schedule - start_at_step and end_at_step slice it rather than restart it. guidance_schedule and the LM-code handoff also read the full schedule, which is what stops a split run from redoing the guidance ramp in stage two.

The two-stage recipe

Same sampler, scheduler, shift and noise_seed in both nodes, steps=50:

| Stage | add_noise | start_at_step | end_at_step | return_with_leftover_noise | |-------|-------------|-----------------|---------------|------------------------------| | First | enable | 0 | 25 | enable | | Final | disable | 25 | 50 | disable |

Connect the first node's latent output to the second node's latent_or_audio, and hang the VAE off the final stage only. An unfinished range - one that still ends above zero noise, or one that executes no steps at all - skips audio decoding even with a VAE plugged in, so a mid-pipeline decode just isn't a thing you should expect.

Outputs

model, vae, positive and negative pass through for chaining. latent is the raw diffusion result - the one you chain into the next stage. audio is decoded audio, and it only appears when a VAE is connected; without one the node prints a note to console and hands back nothing to wire.

The pack's own knobs, and why they're there

This node also carries the newer widgets the author added while chasing ACE-Step 1.5 SFT quality, and two of them are on by default for good reason.

timbre_cls_token=true restores a learned summary token that the timbre encoder in this ComfyUI core drops - it gets prepended before audio and RoPE positions, matching the official XL implementation. schedule_precision=float32 is the other: building the ace_step schedule in BF16 loses precision before it's converted, and CPU checks at shift=3 found 3, 13 and 35 repeated-sigma intervals for 50, 100 and 200 steps - points where the integrator simply doesn't advance. Float32 is the default now; model reproduces the old behaviour if you want to A/B it.

Then the experimental ones: guidance_rescale (0 by default; try 0.25 and 0.5 with identical seeds - it moves the guided velocity's standard deviation toward the conditional prediction's, at no extra model evaluations), apg_norm_rms (0 keeps the fixed L2 cap from apg_norm_threshold; positive values replace it with an RMS cap independent of duration - not an audio volume control), and guidance_mode=apg_denoised_joint, which projects clean predictions across channels and time together. The author is refreshingly blunt that musical improvement from these needs listening comparisons, not test suites.

Install

Manager search ComfyUI-AceStep_SFT, or:

cd ComfyUI/custom_nodes
git clone https://github.com/jeankassio/ComfyUI-AceStep_SFT.git

Install requirements into ComfyUI's Python (transformers>=4.52.0, torchaudio, librosa, soundfile, numpy, pyyaml), drop in the model files (acestep_v1.5_sft.safetensors or the merged SFT+Turbo build, the Qwen encoders, ace_1.5_vae.safetensors), then restart. You need a ComfyUI with native ACE-Step 1.5 support; the pack layers on the core's ACE modules.

Common issues

jkass_quality is missing, or you have two of them. It's bundled from the author's JK-AceStep-Nodes pack, and the local copy doesn't touch the global sampler registry - so you get exactly one entry whether or not JK is installed separately. If the sampler dropdown looks stale, restart; the pack appends new widgets after existing ones and saved workflows keep their old selections.

An old workflow won't run. ace_step_euler was removed - select jkass_quality and re-save. Values from dropped widgets (enable_normalization, velocity_ema_factor and friends) don't come back; output peak protection stays.

Console warns that your duration doesn't match the LM encoding. Keep TextEncode's and Generate's duration equal, or set Generate's to 0 to inherit.

The continuation doesn't sound like an unbroken run. Sometimes it won't. APG momentum, multistep solver history and SDE noise state may reset at the node boundary; deterministic standard_cfg with the Heun-based jkass_quality survives continuation best. Use stages when you need a decision point, not as a ritual.