VRGDG Music3 Tuning Presets
Pick a Behavior, Not a Bunch of Numbers
- ar_cfg_scale
- ar_top_k
- sampler_cfg
- steps
- conditioning_strength
- notes
MiniMax Music3 has more knobs than most people ever turn: two on the text encoder, two on the sampler, one hidden in the conditioning. VRGDG Music3 Tuning Presets collapses all of them into a single dropdown that picks a behavior - "I want it to follow the lyrics strictly" or "I want it looser and weirder" - instead of a pile of numbers you'd have to guess. Twenty presets, each of which outputs the five values Music3's sampler chain actually consumes.
It's one of the Music3 helpers in the VRGameDevGirl pack (vrgamegirl19/comfyui-vrgamedevgirl), the companion node set to the pack's LTX 2.3 / MiniMax H3 music-video builder. Where the Caption Builder decides what style to request, this node decides how Music3 interprets it - the two are deliberately independent, which is exactly why it's useful.
How it works
Music3 is two-stage: an autoregressive stage writes a hidden "acoustic plan" from your caption, then a diffusion stage renders audio from that plan. The five outputs map onto those stages:
ar_cfg_scale(FLOAT) andar_top_k(INT) → Music3 Text Encode inputs. These steer the AR stage: higher cfg follows the caption/lyrics more strongly, lower top_k is more constrained and repeatable.sampler_cfg(FLOAT) andsteps(INT) → KSampler inputs. These steer the diffusion stage: cfg is denoiser guidance between the plan and zeroed conditioning, steps are render refinement.conditioning_strength(FLOAT) → the pack's Conditioning Strength node, which scales how hard the DiT leans on the acoustic plan.
So a preset is really a coordinated set of values across all three nodes. The tooltip text on the dropdown doubles as a manual: "Balanced / Built-in Baseline" is the reference point you should judge everything against; "Fast Draft" uses only 20 steps for quickly testing lyrics, captions, and seeds; "Strict Lyrics" is the strongest lyric guidance; "Maximum Arrangement Variation" is the loosest and most exploratory (and the author flags its drift risk in the same breath).
What you actually do with it
Pick "Balanced / Built-in Baseline" first. Generate a track. Listen. Then pick the preset that names the problem you heard - "didn't follow the lyrics" → "Lyrics + Structure Focus" or "Strict Lyrics"; "too stiff" → "Creative Variations" or "Creative Melody"; "harsh artifacts" → "Conservative / Stable" or "Stable / Low Artifact." Treat the presets as opinions, not verdicts: the notes output tells you what each one does, and you can always override individual outputs by wiring a constant in front of them.
For iterating on lyrics or captions, Fast Draft is the one you'll reach for constantly - 20 steps is plenty to hear whether a new verse lands, and it makes A/B loops tolerable.
Installing it
Same one-time install as the rest of the pack. ComfyUI Manager → Install Custom Nodes → search vrgamedev, or:
cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git
then pip install -r requirements.txt, restart, hard-refresh. It's a heavy pack - requirements include TTS/LLM wheels (voxcpm, llama-cpp-python) you won't touch for Music3. On Windows portable, install Cython and scikit-build-core first and prefer Python 3.12.
Troubleshooting
A preset that "doesn't work" is usually a wiring problem: check that ar_cfg_scale/ar_top_k actually reach the Text Encode and sampler_cfg/steps reach the KSampler, and that conditioning_strength goes through a Conditioning Strength node rather than straight into the sampler. The classic silent failure is leaving the Text Encode's own cfg/top_k widgets at their defaults, which can override or fight the preset's values. And remember presets change behavior, not composition - if two "tighter lyric" presets both drift, the caption itself is the problem, and that's the Caption Builder's department.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| preset | COMBO | Caption presets choose WHAT style to request. Tuning presets choose HOW Music3 interprets it. Balanced / Built-in Baseline: Neutral reference using Music3's familiar starting values. Expect balanced adherence and variation without pushing any stage; use this before judging another preset. Balanced Quality: Adds rendering steps and slightly tightens planning. Tends to sound a little more finished and deliberate while remaining broadly balanced. Raw Live Rock: Firm planning and stronger diffusion guidance with moderate plan emphasis. Tends toward decisive, energetic, punchy results; may become slightly rougher than baseline. Lyrics + Structure Focus: Tighter AR choices and stronger text guidance. Tends to follow words and section labels more closely, with somewhat less melodic and arrangement freedom. Creative Variations: Looser guidance, high planning diversity and gentler rendering. Tends to explore different melodies, phrasing and arrangements; lyrics or structure may wander more. Intimate + Dry: Strong caption guidance, constrained planning and restrained diffusion. Tends toward focused, simpler and less explosive results that leave more room around the lead. Aggressive + Dense: Higher diversity, rendering guidance, steps and plan strength. Tends toward thicker, more forceful and busier results; carries greater harshness or artifact risk. Conservative / Stable: Constrained planning with lower diffusion and plan strength. Tends to produce steadier, cleaner and less adventurous results when stronger presets become unstable. Strict Lyrics: The strongest lyric guidance with the narrowest token choice. Tends to prioritize word and line adherence; may sound more rigid or less inventive. Strong Song Structure: High AR guidance, low diversity, extra steps and stronger plan emphasis. Tends to make verses, choruses and transitions more deliberate; may reduce spontaneity. Creative Melody: Low AR guidance with high token diversity and moderate rendering. Tends to produce more surprising melodies and vocal phrasing while still retaining a usable structure. Maximum Arrangement Variation: The loosest and most exploratory preset. Tends to create the largest changes in instrumentation, pacing and section treatment; has the highest drift risk. Stable / Low Artifact: Constrained planning, restrained diffusion and reduced plan strength with extra steps. Tends to favor smooth, reliable rendering over impact or novelty. Strong Acoustic Plan: Emphasizes a fairly focused hidden music plan during diffusion. Tends to sound decisive and tightly committed to the generated arrangement; excessive emphasis can sound harsh. Loose Acoustic Plan: High planning diversity with low text, diffusion and plan strength. Tends to feel freer, softer and less locked-in; melody, lyrics and sections may drift. Detailed Rendering: Keeps baseline composition behavior but raises diffusion steps substantially. Tends to refine texture and separation rather than rewrite the composition; slower to generate. Long-Song Stability: Tighter planning, stronger text guidance and more rendering steps. Intended to help longer songs remain coherent across sections; may repeat ideas or feel less spontaneous. Fast Draft: Uses only 20 rendering steps with otherwise moderate controls. Useful for quickly testing lyrics, captions and seeds; expect less refinement than a final render. High-Energy / Dense: Strong diffusion, more steps, elevated diversity and plan emphasis. Tends toward louder-feeling, layered and high-impact arrangements; increased congestion or artifact risk. Soft / Restrained: Focused AR planning with gentle diffusion and slightly reduced plan strength. Tends toward calmer dynamics, lighter density and less aggressive rendering. |
Outputs (6)
| Name | Type | Description |
|---|---|---|
| ar_cfg_scale | FLOAT | Connect to Music3 Text Encode cfg_scale. Higher follows caption/lyrics more strongly while generating the AR plan. |
| ar_top_k | INT | Connect to Music3 Text Encode top_k. Lower is more constrained/repeatable; higher explores more token alternatives. |
| sampler_cfg | FLOAT | Connect to KSampler cfg. Controls denoiser guidance between the acoustic plan and zeroed conditioning. |
| steps | INT | Connect to KSampler steps. More steps can refine detail but do not guarantee better composition. |
| conditioning_strength | FLOAT | Connect through Music3 Conditioning Strength after Text Encode. Scales the acoustic plan presented to the DiT. |
| notes | STRING | Preset-specific notes and explanations. |