VRGDG Music3 Conditioning Blend (Experimental)

If you've ever A/B'd two captions in MiniMax Music3 and wished the good one was a little of both, this is the node you were looking for. VRGDG Music3 Conditioning Blend (Experimental) takes the conditioning from two separate Music3 Text Encode runs and interpolates between them, then hands the result to the sampler. Where the caption builder and tuning presets help you design one song, this one helps you find the song that lives halfway between two takes on the same lyrics.

It comes from the VRGameDevGirl pack (vrgamegirl19/comfyui-vrgamedevgirl), a big utility collection built around AI music-video production - the flagship is a full video builder for LTX 2.3 and MiniMax H3. These Music3 nodes are the "make the soundtrack sane" part of that picture.

How the blend works

Music3 is a two-stage model. An autoregressive stage writes a hidden "acoustic plan" token sequence from your caption, then a diffusion stage renders audio from that plan. The thing this node blends is those plans, not the audio or even the text. That's a meaningful difference: a blend in plan space is cheap, deterministic, and produces a single coherent render rather than two songs glued together.

Two of the four inputs are the plans themselves: conditioning_a (your primary caption) and conditioning_b (an alternative). The blend_b slider does the work - 0 gives you only A, 1 only B, and the default 0.35 leans toward A while letting B flavor it. The author's own tooltip spells out the golden rule for useful results: A and B should use the same lyrics, duration, and preferably the same seed. Blend two songs with different lyrics and you don't get a mashup, you get mush.

length_mode exists because the two plans can end at different times. crop to shorter is the safe default; pad shorter with zeros preserves the longer requested duration but can leave a dead tail. Start with crop.

Inputs and outputs that matter

• conditioning_a / conditioning_b - the two CONDITIONING plans from Music3 Text Encode.
• blend_b - the only knob you'll touch. 0–1, step 0.05.
• length_mode - crop vs. pad for mismatched plan lengths.

Outputs: conditioning (wire it into your KSampler's conditioning input, exactly where a single Text Encode's output would go) and blend_report - a string that tells you what you actually blended, handy for logging which mix produced a keeper.

Installing it

Install once, it serves every node in this pack. Easiest via ComfyUI Manager → Install Custom Nodes → search vrgamedev, or paste the repo URL. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

then install requirements.txt with the same Python that runs ComfyUI, and restart + hard-refresh the browser page. Note this pack is heavy: installing it drags in TTS and LLM dependencies (voxcpm, llama-cpp-python) you don't need for the Music3 nodes. On the Windows portable build the README has you install Cython and scikit-build-core first, and Python 3.12 is the safer bet than 3.13.

Troubleshooting

The "Experimental" in the name isn't decoration - the tooltip warns some combinations become incoherent, so blend conservatively (0.2–0.5) and expect to try a couple of settings. If the result sounds like neither A nor B, you've violated the same-lyrics rule or blended plans with very different lengths; re-run both through the same seed. And if you're about to blend two captions that only differ by one word, you've found the other use for this node - it's a perfectly good tool for confirming a change doesn't matter, not just for chasing a middle ground.