fix(stg): drop unsupported momentum_buffer kwarg from APG call (#492)

[Tai An (Anai-Guo)]

Fixes #492

Problem

STGGuiderAdvanced.predict_noise calls apg() with a momentum_buffer=None
keyword that apg() does not accept:

# stg.py:55
def apg(
    noise_pred_pos: torch.Tensor,
    noise_pred_neg: torch.Tensor,
    cfg_scale: float,
    eta: float = 1.0,
    norm_threshold: float = 0.0,
):

# stg.py:459 -- reached on every sampling step when apply_apg=True
if self.apply_apg:
    stg_result = apg(
        stg_result,
        noise_pred_neg,
        cfg_scale=self.apg_cfg_scale,
        momentum_buffer=None,   # <-- not a parameter of apg()
        eta=self.eta,
        norm_threshold=self.norm_threshold,
    )

apply_apg is a node input on LTXVApplySTGAdvanced (stg.py:626), so turning
it on makes sampling fail immediately with

TypeError: apg() got an unexpected keyword argument 'momentum_buffer'

which is exactly what #492 reports.

Why removing it is the right fix

• apg()'s body never references momentum at all, so the argument was inert
  even before it started raising — passing None could not have had an effect.
• The sibling call site, STGGuiderAPG.predict_noise at stg.py:857, already
  calls apg(noise_pred_pos, noise_pred_neg, self.cfg_scale, self.eta, self.norm_threshold) with no momentum argument. Of the two apg() call sites
  in the file, only this one disagrees with the signature.
• It is also the workaround the reporter of apg() got unexpected keyword argument 'momentum_buffer' when apply_apg=True #492 confirmed works.

Restoring real momentum support would be a behaviour change (MomentumBuffer is
constructed in STGGuiderAPG.__init__ and reset on timestep increase, but is
never threaded into apg()), so I left that alone — happy to open a separate
issue for it if you'd like it tracked.

Verification

No ComfyUI runtime needed: project() and apg() were lifted out of stg.py
with ast and executed as-is against real tensors.

apg signature: (noise_pred_pos, noise_pred_neg, cfg_scale, eta=1.0, norm_threshold=0.0)

[1] STGGuiderAdvanced.predict_noise, stg.py:460 -- as shipped
    TypeError: apg() got an unexpected keyword argument 'momentum_buffer'
    matches issue #492 verbatim: True

[2] stg.py:460 with `momentum_buffer=None,` removed (this PR)
    returns (1, 4, 8, 8) torch.float32

[3] STGGuiderAPG.predict_noise, stg.py:857 -- sibling, unmodified
    returns (1, 4, 8, 8)
    fixed call == sibling call: True

[4] does apg's body reference momentum at all?
    'momentum' appears in apg() body: False

So the unpatched path reproduces the reported error verbatim, and the patched
call is bit-identical to the already-working STGGuiderAPG path.

Notes

• 1 line removed, no behaviour change for apply_apg=False.
• black 24.4.2 (the pinned .pre-commit-config.yaml version) leaves the call
  formatting unchanged — the magic trailing comma keeps it exploded.
• PR fix: remove unsupported APG momentum buffer argument #500 proposed the same one-line change but was self-closed by its author on
  2026-08-26 without review, alongside two of their unrelated PRs; apg() got unexpected keyword argument 'momentum_buffer' when apply_apg=True #492 is still
  open and the bug is still on master. Reopening that fix here with runtime
  evidence. Credit to omribz156 for spotting it first.

🤖 Generated with Claude Code