Hugging Face's logo

ntc-ai
/
yue2-concept-sliders

Text-to-Audio
English
yue2
lora
music-generation
comfyui
Model card Files
xet
 Community
2

ComfyUI node: voice sliders work, genre sliders (metal) not audible - tips?

#1
by becausereasons - opened
Discussion
becausereasons

Hi, thanks for releasing these. We installed the ComfyUI node and the particle weights and everything runs without errors, but we are not noticing any audible difference from the sliders. Hoping you can tell us what we are doing wrong, or whether this is expected.

Setup

  • Windows 11, RTX 5090, ComfyUI master 0d901722 (includes your verified commit 387f98aa), torch 2.10.0+cu130, launched with --use-sage-attention
  • comfyui/ntc_yue2_sliders node from this repo (copied into custom_nodes), base yue2_3b_bf16.safetensors from Comfy-Org/YuE2
  • weights/particle-1200-v1/*_step1200.safetensors, all 16 show up in the node and load

Test (your bundled workflow.json, unchanged apart from duration)

  • release caption + lyrics from the workflow, seed 1709, empty ABC (so off mode), cfg_scale 1.01, max_duration 60
  • KSampler 16 steps dpm_2 / sgm_uniform, cfg 1.0, VAEDecodeAudio, FLAC out
  • three runs: slider off, male_step1200 at strength 1.0, metal_step1200 at strength 1.0

Result

  • off: 36 s song. male ON: 60 s song. metal ON: 59 s song. Waveform correlation between off and either ON run is ~0, so the node is clearly changing the plan (the _generate object patch is active).
  • But the lead voice stays female on the male run, and the metal run is the same synth-pop with no distorted guitars or aggressive drums. An audio-LLM judge (Gemini) agrees on both: FEMALE / FEMALE / FEMALE, metal 0/10 on all three.
  • Same outcome earlier on ComfyUI v0.36.0 (before your commit) with abc empty and strength 1.0.

Questions

  1. Is a 60 s render in off mode the right regime, or do the sliders only show clearly in the ~20 s / 500-token setting your samples use?
  2. Any known gap between the Space runtime (native YuE2 pipeline) and the ComfyUI node on a real GPU run, e.g. dtype (bf16 base, fp32 adapter), sage attention, or the AR CUDA-graph path?
  3. Recommended strength / temperature / top_p to make the effect obvious, and is 1.0 really the intended endpoint in Comfy?
  4. Would you expect the distilled rank-8 _comfyui LoRAs (standard Load LoRA, MODEL 0 / CLIP 1) to behave better or worse than the particle node?

Happy to run any diagnostic you suggest and report back. Thanks!
becausereasons changed discussion title from ComfyUI node runs but we hear no steering (male / metal sliders) - tips? to ComfyUI node: voice sliders work, genre sliders (metal) not audible - tips?
becausereasons

Correction: after listening ourselves, the male slider DOES work: the male-ON render has a clearly male lead where the off render is female. Our automated judge (Gemini audio) mislabelled the voice, so please disregard that part of the report. Apologies for the noise.

So the open question narrows to the genre sliders: the metal run at strength 1.0 did not sound metal to us on that prompt (still synth-pop). Any tips on which genre sliders show most clearly, and on strength / duration for those, would still be appreciated.
ntc-ai
 Owner
•
edited about 5 hours ago

Thanks for the clear writeup — and for the correction that male does work. That helps.

Important: native particle weights do not work with stock ComfyUI (including Load LoRA). They only run through our custom node: ntc_yue2_sliders (zip).

This repo ships different formats — use the matching loader:

Weight type Typical path / filename Load with Do not use
Native particles (routed AR adapters) weights/particle-1200-v1/*_step1200.safetensors YuE2 Concept Slider (ntc-ai) custom node only Standard Load LoRA — particles will not apply correctly in Comfy without the plugin
Distilled Comfy LoRAs distilled-rank8-v1/comfyui/*_comfyui.safetensors Comfy Load LoRA on the YuE2 CLIP/TE The particle slider node (expects particle meta + bridges)
Native / non-Comfy exports (if you grabbed those) non-comfyui distilled or training exports Native / Diffusers path only Dropping them on Comfy Load LoRA or the particle node

You already described the particle node + *_step1200 setup, so if that’s exact and there’s no extra Load LoRA on the graph, you’re on the intended path for particles.

If metal still isn’t landing after that checklist: try the distilled Comfy metal file via Load LoRA (separate path), and avoid stacking both particle node + Load LoRA on the same run.

Happy to dig further if you can paste which exact filenames are in the node / LoRA slot.

Sign up or log in to comment