[GH-ISSUE #23] Bug: Create Voice does not recompute when audio input changes (text unchanged) — likely broken IS_CHANGED; also always overwrites output file with no versioning #21

New issue

Open

opened 2026-10-30 21:03:35 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-30 21:03:35 +00:00

Owner

Originally created by @cdshaw on GitHub (Jul 24, 2026).
Original GitHub issue: https://github.com/1038lab/ComfyUI-QwenTTS/issues/23

Bug: Create Voice does not recompute when audio input changes (text unchanged) — likely broken IS_CHANGED; also always overwrites output file with no versioning

Bug 1: Stale caching when audio changes but text stays the same

Observed behavior: Swapping the audio input to a new reference clip while leaving text unchanged does not trigger recomputation — the node appears to reuse a cached result instead of processing the new audio.

Likely root cause: every run of this node (regardless of inputs) logs:

WARNING: _.IS_CHANGED() got an unexpected keyword argument 'format'

IS_CHANGED() is what ComfyUI calls to decide whether a node's inputs have changed and it needs to recompute vs. reuse a cached output. If this method is throwing due to a signature mismatch with ComfyUI's current calling convention, it would explain why input changes aren't being detected — the node may be silently falling back to "unchanged" on every run.

Steps to reproduce:

Run Create Voice once with audio = clip A, text = "some transcript".
Change audio to a different clip B, leave text identical.
Run again.
Observed: output is identical to the first run (or executes suspiciously fast / clearly doesn't reprocess clip B).

Expected behavior: Any change to audio (or any other input) should trigger a full recompute, matching the IS_CHANGED warning shown in the console — that warning itself suggests the detection logic needs fixing to match ComfyUI's current API signature.

Bug 2: Output .pt file is always overwritten, no versioning

Observed behavior: Running Create Voice multiple times with the same name value silently overwrites the previous .pt file at ComfyUI/output/qwen3-tts_voices/<name>.pt — no warning, no auto-incrementing filename, no versioned copy kept.

Expected behavior: Either:

Auto-increment the filename when a file with the same name already exists (e.g. Josie.pt → Josie_2.pt), similar to how ComfyUI's built-in image/audio save nodes avoid collisions, or
At minimum, log a warning when an existing .pt file is about to be overwritten, so it's not silent.

Workaround in use: manually varying the name field (e.g. appending a date or version number) before each save.

Environment

ComfyUI-QwenTTS installed via git clone into a Stability Matrix–managed ComfyUI instance
ComfyUI 0.28.0, Windows, RTX 4070 (12GB VRAM)
Model: Base 1.7B, cuda:0, bf16

This is a follow-up to a previously filed issue on this repo regarding save_path silently failing to save when pointed at a non-existent subfolder.

Originally created by @cdshaw on GitHub (Jul 24, 2026). Original GitHub issue: https://github.com/1038lab/ComfyUI-QwenTTS/issues/23 ## Bug: Create Voice does not recompute when audio input changes (text unchanged) — likely broken IS_CHANGED; also always overwrites output file with no versioning ### Bug 1: Stale caching when audio changes but text stays the same **Observed behavior:** Swapping the `audio` input to a new reference clip while leaving `text` unchanged does not trigger recomputation — the node appears to reuse a cached result instead of processing the new audio. **Likely root cause:** every run of this node (regardless of inputs) logs: ``` WARNING: _.IS_CHANGED() got an unexpected keyword argument 'format' ``` `IS_CHANGED()` is what ComfyUI calls to decide whether a node's inputs have changed and it needs to recompute vs. reuse a cached output. If this method is throwing due to a signature mismatch with ComfyUI's current calling convention, it would explain why input changes aren't being detected — the node may be silently falling back to "unchanged" on every run. **Steps to reproduce:** 1. Run Create Voice once with `audio` = clip A, `text` = "some transcript". 2. Change `audio` to a different clip B, leave `text` identical. 3. Run again. 4. Observed: output is identical to the first run (or executes suspiciously fast / clearly doesn't reprocess clip B). **Expected behavior:** Any change to `audio` (or any other input) should trigger a full recompute, matching the `IS_CHANGED` warning shown in the console — that warning itself suggests the detection logic needs fixing to match ComfyUI's current API signature. ### Bug 2: Output .pt file is always overwritten, no versioning **Observed behavior:** Running Create Voice multiple times with the same `name` value silently overwrites the previous `.pt` file at `ComfyUI/output/qwen3-tts_voices/<name>.pt` — no warning, no auto-incrementing filename, no versioned copy kept. **Expected behavior:** Either: - Auto-increment the filename when a file with the same name already exists (e.g. `Josie.pt` → `Josie_2.pt`), similar to how ComfyUI's built-in image/audio save nodes avoid collisions, or - At minimum, log a warning when an existing `.pt` file is about to be overwritten, so it's not silent. **Workaround in use:** manually varying the `name` field (e.g. appending a date or version number) before each save. ### Environment - ComfyUI-QwenTTS installed via `git clone` into a Stability Matrix–managed ComfyUI instance - ComfyUI 0.28.0, Windows, RTX 4070 (12GB VRAM) - Model: Base 1.7B, cuda:0, bf16 ### Related This is a follow-up to a previously filed issue on this repo regarding `save_path` silently failing to save when pointed at a non-existent subfolder.

bsakaguchi referenced this issue

2026-10-30 21:03:38 +00:00

[PR #21] Add compatibility for Transformers 5.x #22