ComfyUI Node
AceStep 1.5 SFT Generate Advanced
AceStep generation with KSampler Advanced noise and step controls. Connect an unfinished LATENT to the next stage with add_noise disabled. Native Cover/Repaint tasks use Generate.
AceStep 1.5 SFT Generate Advanced
- model
- positive
- negative
- vae
- latent_or_audio
- model
- vae
- positive
- negative
- latent
- audio
◄add_noiseenable►
◄noise_seed0►
◄steps50►
◄cfg7.0►
◄sampler_namejkass_quality►
◄schedulerace_step►
◄start_at_step0►
◄end_at_step10000►
◄return_with_leftover_noisedisable►
◄duration60.0►
◄infer_methodode►
◄guidance_modeapg►
◄batch_size1►
◄latent_shift0.00►
◄latent_rescale1.00►
◄fade_in_duration0.0►
◄fade_out_duration0.0►
◄use_tiled_vaetrue►
◄unload_models_after_generatefalse►
◄voice_boost0.0►
◄apg_eta0.00►
◄apg_momentum-0.75►
◄apg_norm_threshold2.5►
◄guidance_interval1.00►
◄guidance_schedulenormal►
◄min_guidance_scale3.0►
◄guidance_scale_text-1.0►
◄guidance_scale_lyric-1.0►
◄omega_scale0.00►
◄erg_scale0.00►
◄cfg_interval_start0.00►
◄cfg_interval_end1.00►
◄shift3.0►
◄noise_sourceofficial►
◄timbre_cls_tokentrue►
◄guidance_rescale0.00►
◄schedule_precisionfloat32►
◄apg_norm_rms0.000►
Categoryaudio/AceStep SFT
Inputs (43)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | AceStep 1.5 diffusion model (from Load Diffusion Model or with LoRA applied). | |
| positive | CONDITIONING | Positive conditioning from AceStep 1.5 SFT TextEncode. | |
| negative | CONDITIONING | Negative conditioning from AceStep 1.5 SFT TextEncode. | |
| add_noise | COMBO | enable | Enable for the first stage; disable when continuing a partially sampled LATENT. |
| noise_seed | INT | 00–18446744073709550000 | — |
| steps | INT | 501–200 | Diffusion inference steps. |
| cfg | FLOAT | 7.01–20 | Classifier-free guidance scale. Use 7 for Base/SFT and 1 for Turbo checkpoints. |
| sampler_name | COMBO | jkass_quality | jkass_quality uses the bundled JK-AceStep-Nodes Heun sampler (ODE), with two model evaluations per step except the final zero-sigma step. |
| scheduler | COMBO | ace_step | ace_step uses Gradio's shifted linear timesteps and CFG timestep bounds. Other schedulers retain ComfyUI step-fraction intervals. |
| start_at_step | INT | 00–10000 | Start index in the full schedule. For continuation, match the previous stage's end_at_step. |
| end_at_step | INT | 100000–10000 | Stop index in the full schedule. Values at or above steps run through the final step. |
| return_with_leftover_noise | COMBO | disable | Enable to keep an unfinished LATENT for another stage; AUDIO decoding is skipped until finished. Disable to force the endpoint to zero noise. |
| duration | FLOAT | 60.00–600 | Duration in seconds. Set to 0 for auto from latent_or_audio. |
| infer_method | COMBO | ode | ode keeps the selected sampler. sde remaps supported ComfyUI Euler/Heun choices to SDE variants; jkass_quality remains ODE. |
| guidance_mode | COMBO | apg | apg follows ACE-Step velocity guidance. apg_denoised projects clean latents per channel; apg_denoised_joint projects channels and time together. ADG = Angle-based Dynamic Guidance. standard_cfg = normal CFG. |
| vaeopt | VAE | VAE for decoding latents to audio. Audio output requires this. | |
| latent_or_audioopt | AUDIO,LATENT | Source audio or LATENT. To continue an unfinished stage, connect its LATENT and disable add_noise. | |
| batch_sizeopt | INT | 11–16 | Number of audios to generate in parallel. Native task context is repeated for each variation. |
| latent_shiftopt | FLOAT | 0.00-0.2–0.2 | Additive shift on latents before VAE decode. |
| latent_rescaleopt | FLOAT | 1.000.5–1.5 | Multiplicative scale on latents before VAE decode. |
| fade_in_durationopt | FLOAT | 0.00–10 | — |
| fade_out_durationopt | FLOAT | 0.00–10 | — |
| use_tiled_vaeopt | BOOLEAN | true | Use context-padded chunks and discard boundary predictions for 1D audio VAE decoding, matching ACE-Step's overlap-discard policy. |
| unload_models_after_generateopt | BOOLEAN | false | Unload models from memory after generation. |
| voice_boostopt | FLOAT | 0.0-12–12 | Gain in dB for the entire mix before peak protection. Does not isolate or boost only the singer. |
| apg_etaopt | FLOAT | 0.00-10–10 | APG eta: parallel component retention. |
| apg_momentumopt | FLOAT | -0.75-1–1 | APG momentum buffer coefficient. |
| apg_norm_thresholdopt | FLOAT | 2.50–15 | APG update norm cap; 0 disables clipping. Velocity units for apg, clean-latent units for apg_denoised. Joint APG uses this value times sqrt(latent channels) as its aggregate cap (2.5 becomes 20 for 64 channels). |
| guidance_intervalopt | FLOAT | 1.00-1–1 | Guidance interval width. 1.0 matches the official full 0-1 guidance interval; -1 uses cfg_interval_start/end. |
| guidance_scheduleopt | COMBO | normal | normal keeps CFG constant; decay decreases from CFG to min_guidance_scale; ramp_up increases from min_guidance_scale to CFG. Ramps follow the active guidance interval. Applies to standard CFG, APG, ADG and split guidance, including jkass_quality. |
| min_guidance_scaleopt | FLOAT | 3.00–30 | Final CFG for decay, or initial CFG for ramp_up. Set to 1 for a fully conditional endpoint. Unused by normal. |
| guidance_scale_textopt | FLOAT | -1.0-1–30 | Split text guidance. Active when both text and lyric > 1.0. |
| guidance_scale_lyricopt | FLOAT | -1.0-1–30 | Split lyric guidance. Active when both text and lyric > 1.0. |
| omega_scaleopt | FLOAT | 0.00-8–8 | — |
| erg_scaleopt | FLOAT | 0.00-0.9–2 | — |
| cfg_interval_startopt | FLOAT | 0.000–1 | — |
| cfg_interval_endopt | FLOAT | 1.000–1 | — |
| shiftopt | FLOAT | 3.00.1–5 | Timestep schedule shift. ACEStep15 default is 3.0. |
| noise_sourceopt | COMBO | official | official draws noise on the model device in model dtype/BTC layout. comfy preserves ComfyUI CPU noise. Batch variations use seed + index. |
| timbre_cls_tokenopt | BOOLEAN | true | Use the trained timbre summary token omitted by this ComfyUI core. Disable only to compare with the previous generation behavior. |
| guidance_rescaleopt | FLOAT | 0.000–1 | Experimental velocity rescale after CFG/APG/ADG, during sampling. 0 preserves current behavior; compare 0.25 and 0.50 with the same seeds. No extra model evaluations. |
| schedule_precisionopt | COMBO | float32 | float32 avoids rounded or repeated ace_step timesteps with BF16 models. model reproduces the previous schedule precision. Does not change model or noise dtype; other schedulers retain their own precision. |
| apg_norm_rmsopt | FLOAT | 0.0000–1 | Optional APG update RMS cap in the selected prediction space, before projection and CFG. 0 keeps apg_norm_threshold. Positive values replace that fixed cap and keep the per-value budget independent of audio duration. For apg_denoised_joint, compare 0.05 and 0.075. Does not control audio volume; ignored by other guidance modes. |
Outputs (6)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| vae | VAE | — |
| positive | CONDITIONING | — |
| negative | CONDITIONING | — |
| latent | LATENT | — |
| audio | AUDIO | — |