MiniMax Music 3 writes full songs on your own GPU, but it has no “loop” switch. Every track has a start and an end, and if you set it to repeat in a game, a stream or a video, you hear the jump.
I wanted background music that repeats forever without anyone noticing. So I built a small ComfyUI node that finds the best bar-aligned loop inside a Music 3 track and stitches the end back into the start.
This guide gives you the whole thing: the model files, the prompts that make loop-friendly music, the free workflow and node, my test results on an RTX 5080, and fixes for the problems I hit.
TL;DR
- Music 3 has no loop mode. It makes songs with a beginning and an end, and the model card warns the tempo may not match your prompt exactly.
- The fix: prompt for a steady instrumental with no intro or outro, render 75 s, then run my Seamless Audio Loop (FJ) node. It picks an exact 8-bar window and folds the next few milliseconds back over the start.
- My results (RTX 5080 16 GB): lo-fi beat and ambient pad loops both measured a seam ratio of about 1.2 (1.0 = as smooth as the music itself) with no click at the wrap. Each 75 s track took 8 to 9 minutes on the int8 model.
- Two gotchas: Music 3 drifted 2 to 3 BPM off the prompt, so leave
bpmat 0 (auto). It also ended a 75 s request at 30.7 s until I fixed the latent length. - Use WAV, OGG or FLAC for real loops. MP3 adds a short silence at the start that breaks gapless repeat.
New to MiniMax’s models in ComfyUI? Start with our MiniMax H3 ComfyUI guide, the hub for every MiniMax workflow on this site.
Why Music 3 Tracks Don’t Loop on Their Own
MiniMax Music 3 is an open music model that MiniMax released in August 2026. ComfyUI supports it natively through the “MiniMax Music 3 Text to Music” template. It makes 32 kHz stereo songs up to about five minutes long from a structured caption and lyrics.
It was built to write songs, not loops. That causes three problems for looping:
- Songs have shape. The model likes intros, builds and endings, so the last bar rarely sounds like the first.
- Tempo is a suggestion. The model card says the generated tempo, key and structure “may not always match every requested detail exactly”. In my tests, 84 BPM came out near 86, and 70 came out near 72.7.
- No edit modes. The card lists only non-streaming text-to-music. There is no continuation, repaint or loop option to lean on.
So looping has to happen after generation. The trick is to cut in the right place and hide the join.
How the Seamless Loop Works
My node, Seamless Audio Loop (FJ), takes any ComfyUI AUDIO and does four things:
- Measures the real tempo. It builds an onset envelope (where notes and drums hit) and finds the repeating period. You can type the BPM instead, but auto is safer with Music 3.
- Finds the best window. For a loop of N bars, it slides through the track and scores how closely the audio right after the end matches the audio right after the start, using 48 frequency bands.
- Lines up the waveforms. It nudges the length by up to 20 ms so the waves at the seam are in phase, which stops clicks and flams.
- Folds the overhang. It takes the few milliseconds that follow the loop in the original track and crossfades them over the loop’s first few milliseconds (equal-power). The last sample now flows into the first just as the music did originally.
Why not just crossfade the ends? A plain crossfade blends two unrelated moments and smears the beat. Folding the overhang mixes the start with the audio that really followed the end, so the rhythm stays on the grid.
It also prints a seam ratio: how big the spectral jump at the wrap is, compared with the jump at the same spot in the uncut track. About 1.0 means the seam is as smooth as the music itself.
I treat anything up to 1.5 as clean, and I still give each loop one listen on headphones.
Files, Folders and VRAM
These are the official ComfyUI files from Comfy-Org/MiniMax-Music-3. Pick one diffusion model, then grab the text encoder and VAE.
| File | What it is | Size | Folder |
|---|---|---|---|
minimax_music3_dit_int8_convrot.safetensors |
Diffusion model, int8 (what I used) | 2.33 GB | ComfyUI/models/diffusion_models/ |
minimax_music3_dit_fp16.safetensors |
Diffusion model, fp16 (template default) | 4.58 GB | ComfyUI/models/diffusion_models/ |
minimax_music3_text_encoder_pruned_int8_convrot.safetensors |
Text encoder (the 8B planning LLM) | 8.57 GB | ComfyUI/models/text_encoders/ |
minimax_music3_dav.safetensors |
Audio VAE (decoder) | 206.7 MB | ComfyUI/models/vae/ |
The model card says full precision fits under 24 GB, and about 22 GB with automatic CPU offload. It adds that layer-by-layer streaming can fit 8 GB cards, though much slower.
On my 16 GB RTX 5080, the int8 model with the int8 encoder ran fine, and its output came out about the same as fp16 in my comparison. Need help with folders? See the ComfyUI model folders cheat sheet.
Licence: the weights use the MiniMax-Music3 Community License. According to Comfy’s tutorial, companies with under 20 million USD yearly revenue can use it commercially; bigger companies need a commercial licence from MiniMax.
Install the Loop Node
The node is one Python file that uses only numpy and torch, which ComfyUI already has. Nothing to pip install.
- Update ComfyUI. Music 3 needs a recent build; the template appears under Template Library, Audio. Our update guide covers every install type.
- Create the folder
ComfyUI/custom_nodes/ComfyUI-FJ-SeamlessAudioLoop/. - Save the code below as
__init__.pyinside that folder (use the download button). - Restart ComfyUI. Double-click the canvas and search “Seamless Audio Loop”. It sits under
audio/loop.
Show the node code (MIT)
"""ComfyUI-FJ-SeamlessAudioLoop (freakyjolly.com, MIT)
One node that turns a generated track (for example MiniMax Music 3) into a
bar-aligned, seamless loop:
1. estimate the tempo (or use the BPM you prompted for),
2. search for the N-bar window whose end sounds most like its start,
3. nudge the length a few ms so the waveforms line up at the seam,
4. fold the audio that naturally follows the loop back over its first
few milliseconds (equal-power crossfade), so the last sample flows
straight into the first.
No extra pip packages: numpy and torch only (both ship with ComfyUI).
"""
import numpy as np
HOP = 512
NFFT = 2048
def _stft_mag(x):
win = np.hanning(NFFT).astype(np.float32)
n = 1 + max(0, (len(x) - NFFT) // HOP)
idx = np.arange(NFFT)[None, :] + HOP * np.arange(n)[:, None]
frames = x[idx] * win
return np.abs(np.fft.rfft(frames, axis=1)).astype(np.float32) # [frames, bins]
def _bands(mag, n_bands=48):
# log-spaced band energies, log compressed (a cheap mel stand-in)
nb = mag.shape[1]
edges = np.unique(np.geomspace(2, nb - 1, n_bands + 1).astype(int))
out = np.stack([mag[:, a:b].mean(axis=1) for a, b in zip(edges[:-1], edges[1:])], axis=1)
return np.log1p(out * 10.0)
def _onset_env(bands):
flux = np.maximum(0.0, np.diff(bands, axis=0)).sum(axis=1)
flux = np.concatenate([[0.0], flux])
flux -= flux.mean()
return flux / (flux.std() + 1e-9)
def _estimate_bpm(env, sr):
fps = sr / HOP
ac = np.correlate(env, env, mode="full")[len(env) - 1:]
lags = np.arange(len(ac))
lo, hi = int(fps * 60 / 200), int(fps * 60 / 60)
seg = ac[lo:hi]
best = lo + int(np.argmax(seg))
# parabolic refine
if 1 <= best < len(ac) - 1:
a, b, c = ac[best - 1], ac[best], ac[best + 1]
best = best + 0.5 * (a - c) / (a - 2 * b + c + 1e-9)
bpm = 60.0 * fps / best
while bpm < 70:
bpm *= 2
while bpm > 170:
bpm /= 2
return float(bpm)
def find_loop(x_stereo, sr, bpm=0.0, beats_per_bar=4, bars=8,
search_start=4.0, search_end=0.0, window_sec=2.0):
"""x_stereo: float32 [channels, samples]. Returns dict with start, length, bpm, score."""
mono = x_stereo.mean(axis=0).astype(np.float32)
bands = _bands(_stft_mag(mono))
env = _onset_env(bands)
est = _estimate_bpm(env, sr)
use_bpm = float(bpm) if bpm and bpm > 0 else est
L = int(round(bars * beats_per_bar * 60.0 / use_bpm * sr))
Lf = L / HOP
W = max(8, int(window_sec * sr / HOP))
total_f = bands.shape[0]
s0 = int(search_start * sr / HOP)
s1 = total_f - int(np.ceil(Lf)) - W - 2
if search_end and search_end > 0:
s1 = min(s1, int(search_end * sr / HOP))
if s1 <= s0:
raise ValueError(
f"Track too short for {bars} bars at {use_bpm:.1f} BPM "
f"({L / sr:.1f}s loop + {window_sec:.1f}s window after {search_start:.1f}s). "
"Lower bars, lower search_start or render a longer track.")
# normalise band frames for cosine similarity
nb = bands / (np.linalg.norm(bands, axis=1, keepdims=True) + 1e-9)
# 1) refine the loop length: the lag (within +-2.5 % of the nominal bar length)
# at which the track is most similar to itself over the whole search region
lo_f = int(np.floor(Lf * 0.975)); hi_f = int(np.ceil(Lf * 1.025))
region_end = min(total_f, s1 + W + hi_f)
best_lag, best_sim = int(round(Lf)), -1e9
for lag in range(lo_f, hi_f + 1):
a = nb[s0:region_end - lag]
b = nb[s0 + lag:region_end]
if len(a) < W:
continue
sim = float((a * b).sum(axis=1).mean())
if sim > best_sim:
best_sim, best_lag = sim, lag
if bpm and bpm > 0:
# trust the prompted BPM unless the audio clearly disagrees by >1 frame
if abs(best_lag - Lf) <= 1.0:
best_lag = Lf
L = int(round(best_lag * HOP))
Lf = L / HOP
# the refined length can be longer than the nominal one: re-check the search range
s1 = min(s1, total_f - int(np.ceil(Lf)) - W - 2)
if s1 <= s0:
raise ValueError("Track too short for this loop length after tempo refinement. "
"Lower bars or search_start_sec, or render a longer track.")
best = (-1e9, s0)
for s in range(s0, s1):
e = int(round(s + Lf))
a = nb[s:s + W]
b = nb[e:e + W]
cos = float((a * b).sum(axis=1).mean())
pa = nb[max(0, s - W // 2):s]
pb = nb[max(0, e - W // 2):e]
pre = float((pa * pb).sum(axis=1).mean()) if len(pa) and len(pa) == len(pb) else cos
onset_bonus = 0.02 * max(0.0, float(env[s]))
score = 0.6 * cos + 0.4 * pre + onset_bonus
if score > best[0]:
best = (score, s)
score, sf = best
start = sf * HOP
# sample-accurate length nudge (max +-20 ms) by waveform cross-correlation at the seam
span = int(0.02 * sr)
ref_len = int(0.05 * sr)
ref = mono[start:start + ref_len]
best_d, best_c = 0, -1e9
for d in range(-span, span + 1, 4):
seg = mono[start + L + d:start + L + d + ref_len]
if len(seg) < ref_len:
continue
c = float(np.dot(ref, seg) / (np.linalg.norm(ref) * np.linalg.norm(seg) + 1e-9))
if c > best_c:
best_c, best_d = c, d
# fine pass around the coarse best
for d in range(best_d - 4, best_d + 5):
seg = mono[start + L + d:start + L + d + ref_len]
if len(seg) < ref_len:
continue
c = float(np.dot(ref, seg) / (np.linalg.norm(ref) * np.linalg.norm(seg) + 1e-9))
if c > best_c:
best_c, best_d = c, d
L = int(L + best_d)
bpm_from_len = bars * beats_per_bar * 60.0 * sr / L
return {"start": int(start), "length": L, "bpm": bpm_from_len,
"bpm_estimated": est, "score": float(score), "wave_corr": float(best_c)}
def build_loop(x_stereo, start, length, crossfade_samples):
C, N = x_stereo.shape
K = int(min(crossfade_samples, length // 4, max(0, N - (start + length))))
seg = x_stereo[:, start:start + length].copy()
if K > 0:
t = np.linspace(0.0, 1.0, K, dtype=np.float32)
fin, fout = np.sin(t * np.pi / 2), np.cos(t * np.pi / 2)
tail = x_stereo[:, start + length:start + length + K]
seg[:, :K] = seg[:, :K] * fin + tail * fout
return seg, K
def seam_ratio(loop, src, start, sr):
"""Spectral jump across the wrap divided by the jump at the same point in the
original, continuous track. About 1.0 = as smooth as the music itself."""
n = loop.shape[1]
ctx = NFFT * 4
wrap = np.concatenate([loop[:, -ctx:], loop[:, :ctx]], axis=1).mean(axis=0)
a = max(0, start + n - ctx)
orig = src[:, a:a + 2 * ctx].mean(axis=0)
def jump(sig):
b = _bands(_stft_mag(sig))
d = np.linalg.norm(np.diff(b, axis=0), axis=1)
mid = len(d) // 2
return d[max(0, mid - 3):mid + 3].max(), np.median(d)
jw, mw = jump(wrap)
jo, mo = jump(orig)
return float(jw / max(jo, mo, 1e-9))
class FJSeamlessAudioLoop:
@classmethod
def INPUT_TYPES(cls):
return {"required": {
"audio": ("AUDIO",),
"bpm": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 300.0, "step": 0.1,
"tooltip": "0 = estimate from the audio. Otherwise the BPM you put in the caption."}),
"beats_per_bar": ("INT", {"default": 4, "min": 1, "max": 12}),
"bars": ("INT", {"default": 8, "min": 1, "max": 128,
"tooltip": "Loop length in bars. 8 bars at 90 BPM = 21.3 s."}),
"search_start_sec": ("FLOAT", {"default": 6.0, "min": 0.0, "max": 600.0, "step": 0.5,
"tooltip": "Skip the intro: never start the loop before this point."}),
"search_end_sec": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 600.0, "step": 0.5,
"tooltip": "Latest loop start (0 = anywhere). Use it to keep away from the outro."}),
"crossfade_ms": ("INT", {"default": 40, "min": 0, "max": 2000,
"tooltip": "Equal-power fold-over at the seam. 20-80 ms for beats, 200-800 ms for ambient pads."}),
"preview_repeats": ("INT", {"default": 3, "min": 1, "max": 400,
"tooltip": "Second output plays the loop this many times so you can audition the seam (or build a long BGM file)."}),
}}
RETURN_TYPES = ("AUDIO", "AUDIO", "STRING")
RETURN_NAMES = ("loop", "repeated", "report")
FUNCTION = "run"
CATEGORY = "audio/loop"
def run(self, audio, bpm, beats_per_bar, bars, search_start_sec, search_end_sec,
crossfade_ms, preview_repeats):
import torch
wav = audio["waveform"]
sr = int(audio["sample_rate"])
x = wav[0].detach().float().cpu().numpy()
info = find_loop(x, sr, bpm, beats_per_bar, bars, search_start_sec, search_end_sec)
loop, K = build_loop(x, info["start"], info["length"], int(sr * crossfade_ms / 1000))
ratio = seam_ratio(loop, x, info["start"], sr)
rep = np.tile(loop, (1, preview_repeats))
report = (f"start {info['start'] / sr:.3f}s | length {info['length'] / sr:.3f}s "
f"({info['length']} samples) | bars {bars} x {beats_per_bar} | "
f"loop tempo {info['bpm']:.2f} BPM (from audio {info['bpm_estimated']:.2f}) | "
f"match {info['score']:.3f} | wave corr {info['wave_corr']:.3f} | "
f"crossfade {K / sr * 1000:.0f} ms | seam ratio {ratio:.2f} "
f"({'clean' if ratio <= 1.5 else 'check by ear / try more bars or another start'})")
print("[FJ Seamless Audio Loop]", report)
to_t = lambda a: {"waveform": torch.from_numpy(np.ascontiguousarray(a)).unsqueeze(0).to(wav.dtype),
"sample_rate": sr}
return (to_t(loop), to_t(rep), report)
NODE_CLASS_MAPPINGS = {"FJSeamlessAudioLoop": FJSeamlessAudioLoop}
NODE_DISPLAY_NAME_MAPPINGS = {"FJSeamlessAudioLoop": "Seamless Audio Loop (FJ)"}
Tip: if the node doesn’t appear, look in the ComfyUI console for an “IMPORT FAILED” line under ComfyUI-FJ-SeamlessAudioLoop. Usually the file was saved as __init__.py.txt.
Build the Workflow Step by Step
The quickest route is to load my workflow below. It is the official Music 3 template with two changes and the loop node added. Here is what it does, so you can rebuild it yourself.
- Open the template. Template Library, Audio, “MiniMax Music 3 Text to Music”. Let ComfyUI download any missing files.
- Swap to int8 (optional). In the UNETLoader inside the subgraph, pick
minimax_music3_dit_int8_convrot.safetensorson 12 to 16 GB cards. - Fix the length. Open the subgraph and remove the link from the text encoder’s
secondsoutput to EmptyMiniMaxMusic3LatentAudio. Then type 75 into that node’s seconds field. Without this, my first lo-fi track stopped at 30.7 s. - Paste a loop-friendly caption and lyrics from the next section. Set
max_durationto 75 as well. - Add Seamless Audio Loop (FJ) and connect the subgraph’s AUDIO output to its
audioinput. - Save the results. Wire
loopto a SaveAudio node (lossless FLAC, for real use) and to SaveAudioAdvanced (MP3, for sharing). Wirerepeatedto another SaveAudioAdvanced so you can listen to the seam 4 times in a row. - Show the report. Connect
reportto a Preview Any node. It prints the start, length, tempo and seam ratio. - Queue. Files land in
ComfyUI/output/audio/.
Download the workflow and drag it onto the ComfyUI canvas. It needs the loop node installed and the four Music 3 files above.
Show the workflow JSON
{"id":"08882d3e-437d-43da-b498-083de04a60a6","revision":0,"last_node_id":55,"last_link_id":68,"nodes":[{"id":35,"type":"SaveAudioAdvanced","pos":[2150,1250],"size":[750,160],"flags":{},"order":3,"mode":0,"inputs":[{"name":"audio","type":"AUDIO","link":42}],"outputs":[{"name":"audio","type":"AUDIO","links":null}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0"},"widgets_values":["audio/music3_raw","mp3","V0"]},{"id":37,"type":"ac99f841-a3de-4329-9564-953b81cf9e16","pos":[1620,630],"size":[490,540],"flags":{},"order":0,"mode":0,"inputs":[{"label":"tiled_decode","name":"switch","type":"BOOLEAN","widget":{"name":"switch"},"link":null}],"outputs":[{"name":"AUDIO","type":"AUDIO","links":[42,64]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","previewExposures":[]},"widgets_values":["Global Metadata: Lo-fi hip-hop, chillhop instrumental. 84 BPM, A minor, steady and unchanging from start to end, loop-friendly with no intro, no build-up, no breakdown and no ending; the same 4-bar chord cycle repeats the whole time. Study and background listening. Warm dusty bedroom production, vinyl crackle, soft tape saturation.\n\nVocal Details: No vocals. Purely instrumental.\n\nArrangement: Steady boom-bap drum loop with soft kick, brushed snare and swung hi-hats playing the same pattern throughout, round sub bass, mellow Rhodes chords cycling Am7 - Dm9 - G7 - Cmaj7, light jazzy guitar fills, constant vinyl crackle. No intro, no fills that change the energy, no outro, no fade.","[Instrumental]\n\n[Instrumental]\n\n[Instrumental]\n\n[Instrumental]",75,197122968890040,"minimax_music3_dit_int8_convrot.safetensors","minimax_music3_text_encoder_pruned_int8_convrot.safetensors","minimax_music3_dav.safetensors",true]},{"id":50,"type":"FJSeamlessAudioLoop","pos":[2150,630],"size":[420,330],"flags":{},"order":10,"mode":0,"inputs":[{"name":"audio","type":"AUDIO","link":64}],"outputs":[{"name":"loop","type":"AUDIO","links":[65,66]},{"name":"repeated","type":"AUDIO","links":[67]},{"name":"report","type":"STRING","links":[68]}],"properties":{"Node name for S&R":"FJSeamlessAudioLoop"},"widgets_values":[0.0,4,8,6.0,0.0,40,4]},{"id":51,"type":"SaveAudioAdvanced","pos":[2650,560],"size":[420,120],"flags":{},"order":11,"mode":0,"inputs":[{"name":"audio","type":"AUDIO","link":65}],"outputs":[],"properties":{"Node name for S&R":"SaveAudioAdvanced"},"widgets_values":["audio/music3_loop","mp3","V0"]},{"id":52,"type":"SaveAudio","pos":[2650,720],"size":[420,120],"flags":{},"order":12,"mode":0,"inputs":[{"name":"audio","type":"AUDIO","link":66}],"outputs":[],"properties":{"Node name for S&R":"SaveAudio"},"widgets_values":["audio/music3_loop_lossless"]},{"id":53,"type":"SaveAudioAdvanced","pos":[2650,880],"size":[420,120],"flags":{},"order":13,"mode":0,"inputs":[{"name":"audio","type":"AUDIO","link":67}],"outputs":[],"properties":{"Node name for S&R":"SaveAudioAdvanced"},"widgets_values":["audio/music3_loop_x4_preview","mp3","V0"]},{"id":54,"type":"PreviewAny","pos":[2650,1040],"size":[420,160],"flags":{},"order":14,"mode":0,"inputs":[{"name":"source","type":"*","link":68}],"outputs":[],"properties":{"Node name for S&R":"PreviewAny"},"widgets_values":[]},{"id":55,"type":"MarkdownNote","pos":[2150,1000],"size":[440,220],"flags":{},"order":15,"mode":0,"inputs":[],"outputs":[],"properties":{},"widgets_values":["## Seamless loop (freakyjolly.com)\n\n1. Install **ComfyUI-FJ-SeamlessAudioLoop** into `ComfyUI/custom_nodes/`.\n2. Inside the subgraph, the Empty latent is set to a fixed **75 s** (its link from the text encoder's `seconds` output is removed) so Music 3 can't end the song early.\n3. `bpm 0` = measure the real tempo (Music 3 drifts a few BPM off the caption).\n4. Beats: crossfade 20-80 ms. Pads/ambient: 300-600 ms.\n5. Read the report: seam ratio <= 1.5 = clean."]}],"links":[[42,37,0,35,0,"AUDIO"],[64,37,0,50,0,"AUDIO"],[65,50,0,51,0,"AUDIO"],[66,50,0,52,0,"AUDIO"],[67,50,1,53,0,"AUDIO"],[68,50,2,54,0,"STRING"]],"groups":[],"definitions":{"subgraphs":[{"id":"ac99f841-a3de-4329-9564-953b81cf9e16","version":1,"state":{"lastGroupId":0,"lastNodeId":43,"lastLinkId":63,"lastRerouteId":0},"revision":0,"config":{},"name":"Text to Music (MiniMax Music 3)","inputNode":{"id":-10,"bounding":[320,820,128,208]},"outputNode":{"id":-20,"bounding":[3210,380,128,68]},"inputs":[{"id":"9f92b5c5-dadc-4b0f-84d9-b5ed2b3a0842","name":"caption","type":"STRING","linkIds":[45],"pos":[424,844]},{"id":"bb9d2920-ba98-4c51-be7d-4b5736b14407","name":"lyrics","type":"STRING","linkIds":[46],"pos":[424,864]},{"id":"be60d2f1-fc99-4ed6-a057-130a7b83945f","name":"max_duration","type":"FLOAT","linkIds":[47],"pos":[424,884]},{"id":"a7f66596-0413-4906-a08f-475af0521252","name":"seed","type":"INT","linkIds":[54],"pos":[424,904]},{"id":"cbfb7b7b-357f-44ab-80cb-8748020a9880","name":"unet_name","type":"COMBO","linkIds":[55],"pos":[424,924]},{"id":"ec2e20dc-2e7b-4147-922f-5f5f26ac9854","name":"clip_name","type":"COMBO","linkIds":[56],"pos":[424,944]},{"id":"eeefaa94-f342-415e-96e7-eb13dcccbac0","name":"vae_name","type":"COMBO","linkIds":[57],"pos":[424,964]},{"id":"6da74b03-27d0-487e-b8d6-f94018b2ca9a","name":"switch","type":"BOOLEAN","linkIds":[63],"label":"tiled_decode","pos":[424,984]}],"outputs":[{"id":"82bef6fd-2dee-4579-b4ed-7d679d18279f","name":"AUDIO","type":"AUDIO","linkIds":[62],"localized_name":"AUDIO","pos":[3234,404]}],"widgets":[],"nodes":[{"id":6,"type":"UNETLoader","pos":[700,380],"size":[480,90],"flags":{},"order":1,"mode":0,"showAdvanced":false,"inputs":[{"localized_name":"unet_name","name":"unet_name","type":"COMBO","widget":{"name":"unet_name"},"link":55}],"outputs":[{"localized_name":"MODEL","name":"MODEL","type":"MODEL","links":[6]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"UNETLoader","models":[{"name":"minimax_music3_dit_fp16.safetensors","url":"https://huggingface.co/Comfy-Org/MiniMax-Music-3/resolve/main/diffusion_models/minimax_music3_dit_fp16.safetensors","directory":"diffusion_models"}]},"widgets_values":["minimax_music3_dit_int8_convrot.safetensors","default"]},{"id":13,"type":"MiniMaxMusic3TextEncode","pos":[1230,380],"size":[500,750],"flags":{},"order":6,"mode":0,"showAdvanced":true,"inputs":[{"localized_name":"clip","name":"clip","type":"CLIP","link":18},{"localized_name":"caption","name":"caption","type":"STRING","widget":{"name":"caption"},"link":45},{"localized_name":"lyrics","name":"lyrics","type":"STRING","widget":{"name":"lyrics"},"link":46},{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":52},{"localized_name":"max_duration","name":"max_duration","type":"FLOAT","widget":{"name":"max_duration"},"link":47}],"outputs":[{"localized_name":"CONDITIONING","name":"CONDITIONING","type":"CONDITIONING","links":[15,16]},{"localized_name":"seconds","name":"seconds","type":"FLOAT","links":[]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"MiniMaxMusic3TextEncode"},"widgets_values":["","",222,"fixed",60,1.7,50]},{"id":3,"type":"CLIPLoader","pos":[700,550],"size":[480,120],"flags":{},"order":0,"mode":0,"inputs":[{"localized_name":"clip_name","name":"clip_name","type":"COMBO","widget":{"name":"clip_name"},"link":56}],"outputs":[{"localized_name":"CLIP","name":"CLIP","type":"CLIP","links":[18]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"CLIPLoader","models":[{"name":"minimax_music3_text_encoder_pruned_int8_convrot.safetensors","url":"https://huggingface.co/Comfy-Org/MiniMax-Music-3/resolve/main/text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors","directory":"text_encoders"}]},"widgets_values":["minimax_music3_text_encoder_pruned_int8_convrot.safetensors","minimax","default"]},{"id":7,"type":"VAELoader","pos":[700,750],"size":[480,70],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"vae_name","name":"vae_name","type":"COMBO","widget":{"name":"vae_name"},"link":57}],"outputs":[{"localized_name":"VAE","name":"VAE","type":"VAE","links":[13,59]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"VAELoader","models":[{"name":"minimax_music3_dav.safetensors","url":"https://huggingface.co/Comfy-Org/MiniMax-Music-3/resolve/main/vae/minimax_music3_dav.safetensors","directory":"vae"}]},"widgets_values":["minimax_music3_dav.safetensors"]},{"id":10,"type":"ConditioningZeroOut","pos":[1820,540],"size":[230,40],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"conditioning","name":"conditioning","type":"CONDITIONING","link":16}],"outputs":[{"localized_name":"CONDITIONING","name":"CONDITIONING","type":"CONDITIONING","links":[8]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"ConditioningZeroOut"},"widgets_values":[]},{"id":15,"type":"EmptyMiniMaxMusic3LatentAudio","pos":[1780,380],"size":[350,90],"flags":{},"order":7,"mode":0,"inputs":[{"localized_name":"seconds","name":"seconds","type":"FLOAT","widget":{"name":"seconds"},"link":null}],"outputs":[{"localized_name":"LATENT","name":"LATENT","type":"LATENT","links":[21]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"EmptyMiniMaxMusic3LatentAudio"},"widgets_values":[75,1]},{"id":9,"type":"KSampler","pos":[2170,380],"size":[370,270],"flags":{},"order":3,"mode":0,"inputs":[{"localized_name":"model","name":"model","type":"MODEL","link":6},{"localized_name":"positive","name":"positive","type":"CONDITIONING","link":15},{"localized_name":"negative","name":"negative","type":"CONDITIONING","link":8},{"localized_name":"latent_image","name":"latent_image","type":"LATENT","link":21},{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":53}],"outputs":[{"localized_name":"LATENT","name":"LATENT","type":"LATENT","links":[12,58]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"KSampler"},"widgets_values":[1111111112,"fixed",30,1.7,"euler","simple",1]},{"id":12,"type":"VAEDecodeAudio","pos":[2590,380],"size":[230,60],"flags":{},"order":5,"mode":0,"inputs":[{"localized_name":"samples","name":"samples","type":"LATENT","link":12},{"localized_name":"vae","name":"vae","type":"VAE","link":13}],"outputs":[{"localized_name":"AUDIO","name":"AUDIO","type":"AUDIO","links":[60]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"VAEDecodeAudio"},"widgets_values":[]},{"id":38,"type":"SeedNode","pos":[890,970],"size":[270,90],"flags":{},"order":8,"mode":0,"inputs":[{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":54}],"outputs":[{"localized_name":"seed","name":"seed","type":"INT","links":[52,53]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"SeedNode"},"widgets_values":[0,"randomize"]},{"id":42,"type":"VAEDecodeAudioTiled","pos":[2580,500],"size":[280,110],"flags":{},"order":9,"mode":0,"inputs":[{"localized_name":"samples","name":"samples","type":"LATENT","link":58},{"localized_name":"vae","name":"vae","type":"VAE","link":59}],"outputs":[{"localized_name":"AUDIO","name":"AUDIO","type":"AUDIO","links":[61]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"VAEDecodeAudioTiled"},"widgets_values":[1536,64]},{"id":43,"type":"ComfySwitchNode","pos":[2600,690],"size":[270,110],"flags":{},"order":10,"mode":0,"inputs":[{"localized_name":"on_false","name":"on_false","type":"AUDIO","link":60},{"localized_name":"on_true","name":"on_true","type":"AUDIO","link":61},{"localized_name":"switch","name":"switch","type":"BOOLEAN","widget":{"name":"switch"},"link":63}],"outputs":[{"localized_name":"output","name":"output","type":"AUDIO","links":[62]}],"properties":{"cnr_id":"comfy-core","ver":"0.31.0","Node name for S&R":"ComfySwitchNode"},"widgets_values":[false]}],"groups":[],"links":[{"id":18,"origin_id":3,"origin_slot":0,"target_id":13,"target_slot":0,"type":"CLIP"},{"id":16,"origin_id":13,"origin_slot":0,"target_id":10,"target_slot":0,"type":"CONDITIONING"},{"id":6,"origin_id":6,"origin_slot":0,"target_id":9,"target_slot":0,"type":"MODEL"},{"id":15,"origin_id":13,"origin_slot":0,"target_id":9,"target_slot":1,"type":"CONDITIONING"},{"id":8,"origin_id":10,"origin_slot":0,"target_id":9,"target_slot":2,"type":"CONDITIONING"},{"id":21,"origin_id":15,"origin_slot":0,"target_id":9,"target_slot":3,"type":"LATENT"},{"id":12,"origin_id":9,"origin_slot":0,"target_id":12,"target_slot":0,"type":"LATENT"},{"id":13,"origin_id":7,"origin_slot":0,"target_id":12,"target_slot":1,"type":"VAE"},{"id":45,"origin_id":-10,"origin_slot":0,"target_id":13,"target_slot":1,"type":"STRING"},{"id":46,"origin_id":-10,"origin_slot":1,"target_id":13,"target_slot":2,"type":"STRING"},{"id":47,"origin_id":-10,"origin_slot":2,"target_id":13,"target_slot":4,"type":"FLOAT"},{"id":52,"origin_id":38,"origin_slot":0,"target_id":13,"target_slot":3,"type":"INT"},{"id":53,"origin_id":38,"origin_slot":0,"target_id":9,"target_slot":4,"type":"INT"},{"id":54,"origin_id":-10,"origin_slot":3,"target_id":38,"target_slot":0,"type":"INT"},{"id":55,"origin_id":-10,"origin_slot":4,"target_id":6,"target_slot":0,"type":"COMBO"},{"id":56,"origin_id":-10,"origin_slot":5,"target_id":3,"target_slot":0,"type":"COMBO"},{"id":57,"origin_id":-10,"origin_slot":6,"target_id":7,"target_slot":0,"type":"COMBO"},{"id":58,"origin_id":9,"origin_slot":0,"target_id":42,"target_slot":0,"type":"LATENT"},{"id":59,"origin_id":7,"origin_slot":0,"target_id":42,"target_slot":1,"type":"VAE"},{"id":60,"origin_id":12,"origin_slot":0,"target_id":43,"target_slot":0,"type":"AUDIO"},{"id":61,"origin_id":42,"origin_slot":0,"target_id":43,"target_slot":1,"type":"AUDIO"},{"id":62,"origin_id":43,"origin_slot":0,"target_id":-20,"target_slot":0,"type":"AUDIO"},{"id":63,"origin_id":-10,"origin_slot":7,"target_id":43,"target_slot":2,"type":"BOOLEAN"}],"extra":{}}]},"config":{},"extra":{"ds":{"scale":0.7901285864132895,"offset":[-390.5963888708135,-245.41184011189858]},"frontendVersion":"1.48.7","VHS_latentpreview":false,"VHS_latentpreviewrate":0,"VHS_MetadataImage":true,"VHS_KeepIntermediate":true},"version":0.4}Prompts That Loop Well
The caption has three parts: Global Metadata, Vocal Details and Arrangement. For loops, every line should push toward “the same thing, all the way through”.
- Say it out loud: “loop-friendly, no intro, no build-up, no breakdown and no ending”.
- Name a short chord cycle. “The same 4-bar chord cycle repeats the whole time” gives the node many matching windows to pick from.
- No vocals. Sung phrases rarely land on a loop boundary. Write “No vocals. Purely instrumental.”
- Lyrics = tags only. Use
[Instrumental]repeated three or four times, nothing else. - Avoid risers, drops and fades. They are one-way events, and the loop will repeat them.
Prompt 1: lo-fi beat (study, stream, menus)
Global Metadata: Lo-fi hip-hop, chillhop instrumental. 84 BPM, A minor, steady and unchanging from start to end, loop-friendly with no intro, no build-up, no breakdown and no ending; the same 4-bar chord cycle repeats the whole time. Study and background listening. Warm dusty bedroom production, vinyl crackle, soft tape saturation.
Vocal Details: No vocals. Purely instrumental.
Arrangement: Steady boom-bap drum loop with soft kick, brushed snare and swung hi-hats playing the same pattern throughout, round sub bass, mellow Rhodes chords cycling Am7 - Dm9 - G7 - Cmaj7, light jazzy guitar fills, constant vinyl crackle. No intro, no fills that change the energy, no outro, no fade.Lyrics: [Instrumental] four times, each on its own line with a blank line between.
Prompt 2: ambient pad (sleep, focus, game worlds)
Global Metadata: Ambient, drone, meditation instrumental. 70 BPM, D major, calm and constant from start to end, loop-friendly with no intro, no climax and no ending. Background for focus, sleep and video game menus. Wide, soft, airy production with long reverb.
Vocal Details: No vocals. Purely instrumental.
Arrangement: Slowly evolving warm synth pads holding D - A - Bm - G, soft sustained strings, gentle sparse piano notes, faint shimmering textures, no drums, no percussion. The same calm texture throughout, no fade-in, no fade-out.Lyrics: [Instrumental] three times. For more caption ideas, MiniMax offers a caption rewriter skill: npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter.
Loop Node Settings Explained
| Setting | Default | What it does | My advice |
|---|---|---|---|
bpm |
0 | 0 measures the tempo from the audio; any other value uses that BPM | Keep 0 for Music 3, which drifts off the caption |
beats_per_bar |
4 | Beats in one bar | 3 for waltz time, 6 for 6/8 feels |
bars |
8 | Loop length in bars | 8 bars at 86 BPM is about 22 s. Use 16 for less obvious repetition, 4 for short game cues |
search_start_sec |
6 | Never start the loop before this point | Skips the first bars, where the model is still settling in |
search_end_sec |
0 | Latest allowed start (0 = anywhere) | Set it if the track fades or changes near the end |
crossfade_ms |
40 | Length of the fold-over at the seam | 20 to 80 ms for drums; 300 to 600 ms for pads and drones |
preview_repeats |
3 | How many times the repeated output plays the loop |
4 to audition; around 160 turns a 22 s loop into an hour |
The track must be long enough for the loop plus a 2-second check window after search_start_sec. With 8 bars near 86 BPM that is about 30 s, which is why I render 75 s and give the search plenty of room.
My Test Results
I ran both prompts on the second RTX 5080 (16 GB) in my PC with the int8 text encoder: the lo-fi track on the int8 diffusion model and the ambient pad on fp16. Sampling was 30 steps, CFG 1.7, euler / simple, seed 42, as in the template.
| Test | Prompted BPM | Measured BPM | Loop | Crossfade | Seam ratio | Click at wrap | Render time |
|---|---|---|---|---|---|---|---|
| Lo-fi beat | 84 | about 86 | 8 bars | 40 ms | 1.21 | None | 8 to 9 min |
| Ambient pad | 70 | about 72.7 | 8 bars | 400 ms | 1.22 | None | 8 to 9 min |
Listen: the loops and the originals
Each loop below plays 4 times in a row, so the seam passes 3 times. Try to hear where it restarts. The waveform marks every seam with a line.
Lo-fi beat: 8-bar loop (22.6 s), played 4x
Ambient pad: 8-bar loop (26.7 s), played 4x
What I took away from the tests:
- Auto BPM matters. If I had forced 84 BPM, each 8-bar loop would have been around half a second too long, and the beat would drift every repeat.
- int8 is enough. It gave about the same result as fp16 and leaves VRAM free.
- Lock the length. Music 3 can predict a shorter song than you asked for. A fixed latent length stops that.
Warning: the first version of my node crashed with bpm 0 on a short track, because the tempo refinement ran past the end of the audio. The code on this page has the fix. If you copied an earlier version, replace it.
Export for Games, Streams and Video
A loop is only seamless if the player repeats it without a gap. The file format decides that.
| Use | Format | How to loop it |
|---|---|---|
| Unity | WAV or OGG | Tick Loop on the AudioSource |
| Godot | WAV or OGG | Turn on loop in the Import dock, then reimport |
| Website background | OGG or AAC (M4A) | loop attribute on the audio element; test in each browser |
| Stream (OBS) | FLAC or WAV | Media Source with Loop ticked |
| Long YouTube BGM | Any | Raise preview_repeats and save the repeated output as one long file |
To turn the FLAC into OGG or WAV, use ffmpeg: ffmpeg -i music3_loop_lossless_00001_.flac -c:a libvorbis -q:a 6 loop.ogg, or ffmpeg -i music3_loop_lossless_00001_.flac loop.wav.
Why not MP3? MP3 encoders add a short silent padding at the start of the file. Many players don’t remove it, so every repeat has a tiny gap. Keep MP3 for previews and sharing.
Want visuals that loop too? Pair this with our MiniMax H3 seamless loop video guide, which makes still-camera cinemagraph loops in the same ComfyUI.
Other Ways to Loop Music
| Method | Inside ComfyUI | Bar-aligned | Best for |
|---|---|---|---|
| Seamless Audio Loop (FJ), this guide | Yes | Yes | One-click loops right after generation |
| PyMusicLooper (MIT, command line) | No | Finds loop points by similarity | Existing tracks; its split-audio makes intro, loop and outro files, and extend makes long versions |
| Audacity or a DAW by hand | No | If you snap to the grid | Fine control and one-off edits |
| Generate a longer track | Yes | No | Background music where an occasional repeat isn’t a problem |
Looking for other local music models? Our best local text-to-song models and best local background music generators compare YuE2, ACE-Step, Stable Audio and more. The same loop node works on their output too, because it takes any ComfyUI audio.
Troubleshooting
| Problem | Cause | Fix |
|---|---|---|
| Song ends far earlier than max_duration | The text encoder’s seconds output feeds the latent with its own shorter guess |
Unlink it and type a fixed length into EmptyMiniMaxMusic3LatentAudio |
Track too short for 8 bars at ... BPM |
Loop plus check window doesn’t fit after search_start_sec |
Lower bars or search_start_sec, or render longer |
| Beat slips a little on every repeat | Forced BPM doesn’t match the real tempo | Set bpm to 0 |
| Soft “whoosh” or doubled notes at the seam | Crossfade too long for percussive music | Drop crossfade_ms to 20 to 40 |
| Click at the seam on pads | Crossfade too short for slow sounds | Raise crossfade_ms to 300 to 600 |
| Seam ratio above 1.5 | The track changes too much between bars | Try 16 bars, move search_start_sec, or reroll with a stricter “no changes” caption |
| Tiny gap when the file repeats | MP3 padding | Use the FLAC output, or convert it to WAV or OGG |
| Out of memory | fp16 model plus encoder on a small card | Use the int8 files and turn on tiled decode; see ComfyUI errors and fixes |
| Node missing after restart | Wrong folder or file name | It must be custom_nodes/ComfyUI-FJ-SeamlessAudioLoop/__init__.py |
FAQ
Does MiniMax Music 3 have a built-in loop option?
No. It only does text-to-music, so you loop the result afterwards. This guide’s node does that inside ComfyUI.
Can I loop a song with vocals?
You can, but it rarely sounds natural, because sung lines don’t end on loop boundaries. For background use, prompt instrumentals, or loop only an instrumental section.
How long should the loop be?
8 bars (about 20 to 30 s) works for most game and stream music. Use 16 bars when listeners will hear it for a long time, so the repetition is less obvious.
Does the node work with ACE-Step, YuE or Stable Audio?
Yes. It accepts any ComfyUI AUDIO input, so you can connect any music model’s output to it, or a Load Audio node.
Why do you recommend auto BPM when I wrote the BPM in the caption?
Because Music 3 treats tempo as guidance. My 84 and 70 BPM prompts came out near 86 and 72.7, and loops cut at the prompted tempo would drift.
Can I use Music 3 loops commercially?
Under the MiniMax-Music3 Community License, companies with under 20 million USD yearly revenue can, according to Comfy’s tutorial. Larger companies need a commercial licence. Read the licence file on Hugging Face before you ship.
How much VRAM do I need?
16 GB is comfortable with the int8 files. The model card says layer streaming can fit 8 GB cards, at a large speed cost.
How do I make a one-hour version?
Set preview_repeats so the loop length times repeats is about 3,600 s (around 160 for a 22 s loop), and save the repeated output.
Final Thoughts
Music 3 is the strongest open song model I’ve run locally, and a small post-processing step makes it a good loop generator too. Prompt for steadiness, render 75 s, leave the tempo on auto, and let the node choose the cut.
Next, give your loops some visuals with the MiniMax H3 seamless loop video guide, or browse every MiniMax workflow in the MiniMax H3 ComfyUI guide.
Sources
- MiniMaxAI/MiniMax-Music3 model card (architecture, VRAM, tempo note)
- Comfy-Org/MiniMax-Music-3 (ComfyUI files)
- ComfyUI MiniMax Music 3 tutorial (template, caption format, licence summary)
- ComfyUI Wiki: MiniMax Music 3 release
- PyMusicLooper on PyPI
