Professional Loop Transition

MusicGen is terrible at endings. It generates a clip and then just... stops, mid-tone, with a click you can hear from across the room. ProfessionalLoopTransition is the pack's answer for dance-music workflows: it stitches clips together with beat-aligned crossfades so that the tail of one generation flows into the head of the next like a DJ's mix, not a tape splice. It's the most ambitious node in this pack - and the one with the most knobs.

The core idea is simple: keep the last audio you received, and when new audio arrives, blend the end of the old clip into the start of the new one over a transition that's measured in musical beats, not arbitrary seconds.

How it works

It's stateful. The node holds the previous clip per loop_id (a string id you set; different ids = independent chains), detects when its input changes, and crossfades old→new. Before blending it does a few pro moves, all toggleable:

• DC offset removal - strips the constant offset that causes pops and clicks at cut points.
• Phase alignment - cross-correlates the tail of the old clip with the head of the new one and shifts to minimize phase cancellation, so the two don't fight each other during the blend.
• Auto-normalize - balances levels so the transition doesn't jump in volume.
• Zero-crossing detection - finds natural silent points in the waveform for the cleanest possible cut.

The transition_type input picks the algorithm: beat_match (default - beat-aligned crossfade for rhythmic music), phrase_match (same idea tuned for phrase boundaries), zero_cross (cleanest cuts anywhere), and power_match (matches RMS levels first, best when clips vary in loudness). transition_beats (0.25–8) sets how long the transition lasts in beats - one beat at 128 BPM is ~0.47s. The optional crossfade_curve (linear, equal_power, exponential, s_curve) shapes the fade; equal_power holds perceived volume steady, which is why it's the default.

Outputs: looped_audio (the stitched result), conditioning_audio (the original input, passed through for MusicGen continuation), and info (what transition ran, BPM, and status).

The honest framing

This is a transition between successive generations, not a magic "make this one clip loop forever" button. You feed it a stream - run MusicGen a few times with evolving prompts, or alternate SmoothAudioQueue output into it - and it makes the hand-offs seamless. For the first clip there's nothing to transition to, so it just passes through; the info string will say "First loop - no transition". Expect that.

The inputs that matter

• audio, bpm (60–200, must match your actual tempo), transition_type, transition_beats, loop_id - the core five.
• dc_offset_removal, auto_normalize, phase_align - leave on; they're cheap and they work.
• crossfade_curve - only touch when you hear an actual problem (e.g. s_curve if equal_power sounds phasey).

Install

Same pack, ebrinz/ComfyUI-MusicGen-HF - via ComfyUI Manager (search the pack title) or:

cd ComfyUI/custom_nodes
git clone https://github.com/ebrinz/ComfyUI-MusicGen-HF
cd ComfyUI-MusicGen-HF
pip install -r requirements.txt

Restart ComfyUI. No extra model downloads - this is pure DSP on top of whatever audio you feed it.

Troubleshooting

Choppy transitions → increase transition_beats or switch to zero_cross. Volume jumps → power_match plus auto_normalize. Phasey or hollow blends → keep phase_align on and try equal_power or s_curve. And the classic: your BPM input doesn't match the actual tempo of the audio, so the "beat-aligned" transition lands off-grid anyway. Set bpm to the real tempo of what you generated. If transitions never happen at all, check that you're feeding it changing audio - same input means "Loop maintained", and a fresh loop_id means it forgets everything it knew.