Workflows saved with the released 0.5.x must keep loading (ComfyUI restores
widget values by position). prefix_tail_trim_frames had been inserted into
Voice Continue's required inputs, which shifted every later value of a
main-era workflow by one: text_top_p received 50 and the prompt was
rejected. It is an appended optional input again, and
tests/test_widget_layout.py freezes the layouts released on main.
8B (MossTTSDelay):
- Decode Tokens ran into a TypeError: only the 1.7B codec takes
return_stereo, so it is passed only where the signature has it.
- A continuation lost one frame per seam, in the tokens and the audio: the
8B counts its audio_start marker row into start_length. Checked with the
real 8B processor and codec: 25.00 frames instead of 24.00.
- The length cap counted rows, not frames. The delay pattern's fixed rows
(n_vq + 2, n_vq + 1 on Voice Continue) are now added on top.
Models:
- One model at a time. The cached bundle IS the loader output that
ComfyUI's RAM-pressure cache keeps alive, so a switch now empties it;
a node handed such an emptied bundle loads it again, after its own input
checks. Measured live on 16 GB: 1.7B -> 1.7B from model_path stays at one
model's VRAM, and a cached loader output is reloaded.
- tokenizer_path must be a MOSS audio tokenizer; its sample rate is checked
for Hub models too (from the HF cache).
- transformers 5.4.0 / 5.5.0 cannot read either tokenizer config; the loader
says so and names 5.5.1, also on the Windows repo-id retry path.
Tokens and paths:
- Load Tokens only ever uses torch.load(weights_only=True).
- Windows: UNC and device paths (\\host, //, \/, /\, \??\UNC) are refused
in path_override, model_path and tokenizer_path, including behind the
dropdown's output/ prefix -- a shared workflow could otherwise hand a
remote host the user's NTLM hash. A path inside ComfyUI's own input or
output dir stays allowed, even when that dir is a share.
- Token tensors must hold non-negative whole numbers.
- Estimate Tokens counts CJK characters plus the words of the rest, so
mixed text is neither 3x too long nor collapsed to 0.4 s.
- Voice Continue concatenates batch item 0 of previous_audio.
Docs: README and tooltips corrected (4.5B parameters behind the "(1.7B)"
label, 8B is 24 kHz mono, sampling defaults per model, transformers
versions, troubleshooting for 0.6.1-0.6.3 branch workflows, network paths).
Example workflow: the voice reference is ~14 s instead of ~7 s.
Tests: test_model_paths 31, test_generation 14, test_widget_layout 2, every
new check mutation-tested in memory. Live on the local ComfyUI: the example
workflow from origin/main runs end to end, Decode Tokens on the 1.7B, model
switch and reload, network paths refused.
Version 0.6.3 -> 0.6.4.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
