Project

Redrob Tune

~14s Fast · ~23s Pro for 60s audio

Observed wall times from the Jul 28 Fast/Pro preference bake-off on our GPUs (sequential completes, seed 42). Fast is the turbo DiT at 8 steps; Pro is the SFT DiT at 50 steps with APG. Samples below include that run’s winners plus earlier bake-off tracks.

Summary

Music on ACE-Step 1.5 through ComfyUI: Tune Fast and Tune Pro. A 60-second track landed in about 14s on Fast and about 23s on Pro in the Jul 28 preference bake-off. Tune has no standalone release; it ships as the Music tool inside Redrob Studio. Samples below include winners from that run and earlier bake-offs.

My role

Owned Fast/Pro graph decisions (APG, ref-audio removal, step counts) and the Jul 28 preference bake-off. Safety policy for lyrics sits in Studio's prompt gate.

Purpose

Generate usable length audio fast enough to bake off presets the same day, and keep harmful lyric and prompt content out of the product path without pretending “SFW suite” is the safety story.

Architecture

Generation runs on ComfyUI with ACE-Step 1.5 weights. The graph loads a DiT, dual ACE text encoders, and an audio VAE. Style caption, lyrics, BPM, key, language, and duration go into the ACE text encoder. An empty latent audio buffer is sampled with AuraFlow scheduling, decoded by the VAE, and saved as MP3. Studio is the client; preference ratings stay in the eval harness.

Two presets share that graph. Fast is for iteration. Pro turns on Adaptive Projected Guidance because plain DiT CFG on the KSampler path is unreliable for this family. Reference audio was removed: it noised lyric adherence and clouded the latent space, so text and lyrics stay the conditioning signal.

Tune Fast
8 steps · CFG ≈ 1 · no APGacestep_v1.5_xl_turbo_bf16
Tune Pro
50 steps · CFG 1.05 · APGacestep_v1.5_xl_sft_bf16
Graph
DiT + DualCLIP + audio VAE → MP3AuraFlow sampling; SaveAudioMP3

Generation time

The claim that matters for Tune is wall-clock time for a usable track, not a preference p-value against a hosted music API.

On the Jul 28 Fast/Pro preference bake-off (sequential jobs, seed 42, 60-second targets), file timestamps between consecutive completes were about 14 seconds for Fast and about 23 seconds for Pro. Those gaps are observed suite wall times on our GPUs, not a formal latency study with cold starts stripped out.

That is the achievement: roughly a minute of music in about fifteen to twenty-five seconds, depending on preset. Fast enough to run an eleven-prompt Fast/Pro matrix in a few minutes and mark winners the same day.

Content safety

Calling the suite “SFW” is the wrong frame. Ordinary music generation is already non-pornographic by default. The product risk that matters is hate speech, racism, and sexual discrimination in style captions or lyrics. Studio’s LLM prompt gate is where that policy lives; preference bake-offs measure Fast vs Pro quality, not whether the track would pass a vague safe-for-work checklist.

What won

Preference on lyric clarity, style match, and listenability. Earlier bake-off winners (neo-soul, Korean ballad, trap, indie pop, Korean trot, hyperpop, Korean boom bap) stay below. The Jul 28 run added seven more: four Fast (Hindi Bollywood, Atlanta trap, NYC boom bap, Chicago drill) and three Pro (US chart pop, US synth-pop, US pop-rock). Korean ballad, neo-soul, and lo-fi were left unrated on that later pass.

Decisions

  • Replace broken DiT CFG with APG on Pro

    Plain classifier-free guidance on the KSampler path is unreliable for this ACE Step family. Pro uses Adaptive Projected Guidance (CFG 1.05 + APG) on the SFT checkpoint; Fast stays at CFG ≈ 1 on the turbo weights with no APG.

  • Split Fast and Pro on one graph, not one checkpoint

    Eight-step turbo and fifty-step SFT share the same Comfy graph and prompt contract. Iteration speed and polish are different operating points, not different products.

  • Remove reference audio from the generation path

    Ref-audio timbre conditioning noised lyric adherence and clouded the latent space. We took it out so text and lyrics stay the conditioning signal, and so Fast/Pro bake-offs compare the same prompt contract.

  • Treat safety as hate and discrimination policy, not an SFW label

    Ordinary music prompts are already non-pornographic by default. The real product risk is hate speech, racism, and sexual discrimination in lyrics or style text. Studio’s LLM prompt gate is the control; a branded “SFW bake-off” does not substitute for that policy.

Challenges

  • No honest commercial preference study yet

    We did not invent a win rate against hosted music APIs. The measurable claim is wall-clock generation plus the bake-off winners on this page.

  • Not released on its own

    Tune runs as Music inside Studio and is not published as a separate repository or checkpoint. Getting it means cloning Studio and attaching a GPU host.

Impact

  • Same-day Fast/Pro matrices

    Fourteen to twenty-three seconds per sixty-second track made an eleven-prompt matrix practical, so winners could be marked without overnight jobs.

Outputs

  • Audio

    Korean boom bap · Redrob Tune Fast

  • Audio

    Neo-soul night · Redrob Tune Pro

  • Audio

    Korean ballad · Redrob Tune Pro

  • Audio

    Trap rap · Redrob Tune Pro

  • Audio

    Indie pop · Redrob Tune Pro

  • Audio

    Korean trot · Redrob Tune Pro

  • Audio

    Hyperpop · Redrob Tune Pro

  • Audio

    Hindi Bollywood · Redrob Tune Fast

  • Audio

    Atlanta trap · Redrob Tune Fast

  • Audio

    NYC boom bap · Redrob Tune Fast

  • Audio

    Chicago drill · Redrob Tune Fast

  • Audio

    US chart pop · Redrob Tune Pro

  • Audio

    US synth-pop · Redrob Tune Pro

  • Audio

    US pop-rock · Redrob Tune Pro

System

  • Redrob Tune

    ~14-23s

    per 60s track

    Music on ACE Step 1.5 via ComfyUI: Tune Fast (~14s) and Tune Pro (~23s) for a 60s track; APG on Pro; ref audio removed; LLM prompt policy for hate/racism/discrimination; preference winners below