IrodoriTTS-v2 Rescale Config

Of all the nodes in this pack, IrodoriTTS-v2 Rescale Config is the one that looks most like a wall of sliders and does the least for the average user. It's a tuning bundle for two advanced sampling features in the IrodoriTTS-v2 backend, and it outputs a single rescale_config object that the Sampler and Dialogue TTS nodes accept as an optional input. If you never open this node, your workflow still runs perfectly. It's here for the people chasing specific artifacts on the last mile.

So what are the two features?

Temporal score rescaling (rescale_k, rescale_sigma). The backend runs a flow-matching sampler, and these two implement an SNR-based rescale of the score during sampling - the same "temporal score rescale" trick you see in F5-TTS-style flow TTS. It changes how firmly the generation is pushed toward the target in the later, detail-heavy part of the trajectory. rescale_k is the main knob (tooltip says typical range 2.0–6.0); rescale_sigma is a secondary adjustment (typical 0.1–1.0) that only does anything when rescale_k is enabled. If your output sounds unstable, soft, or wobbly on the tail end, a gentle rescale_k around 2–3 is the thing to try.

Speaker K/V scale (speaker_kv_scale, speaker_kv_min_t, speaker_kv_max_layers). This one's about voice fidelity rather than stability: it scales the speaker-conditioning key/value strengths inside the transformer's attention during the late sampling steps, then reverts. The idea is to make the cloned voice grip harder where it matters without over-constraining the whole generation. speaker_kv_scale of 1.0–3.0 is the suggested range; speaker_kv_min_t (default 0.9) is where in the schedule it engages - applied while t >= 0.9, then reverted; speaker_kv_max_layers (-1 = all layers) caps how many transformer layers it touches, mostly a performance guard on big models.

The truncation_factor input is a third, cruder knob that scales the initial noisy latent at the start of sampling (the "truncation" trick for flow models). If you're honest with yourself, you'll probably never touch it.

The defaults tell you everything about the intended usage: every feature is off (-1), and the two helper fields that don't have an off switch (speaker_kv_min_t, speaker_kv_max_layers) carry sane defaults. You opt in to each feature by setting it above zero, per the tooltips. That's the right design for a tuning node - nothing happens until you ask.

The one genuinely useful workflow: your cloned voice keeps drifting toward the reference speaker or losing character on long lines. Wire a Rescale Config into the Sampler or Dialogue node's rescale_config input, set speaker_kv_scale to ~2.0, and hear whether late-stage attention scaling firms it up. If it doesn't, it doesn't - revert and keep the reference-clip advice instead.

It installs with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/kantan-kanto/ComfyUI-DialogueTTS
pip install -r ComfyUI-DialogueTTS/requirements.txt

(restart after, or use ComfyUI Manager → search "ComfyUI-DialogueTTS"). And remember the standing rules for this pack: the main checkpoint sits in models/checkpoints/ and downloads only if you put it there, while the codec and tokenizer fetch themselves on first load. If a dependency error shows up at startup, the transformers<5 pin is the usual suspect when another audio pack is fighting over the same environment.