Skip to content

fix: realtime param/prompt correctness + cosine_weighted interpolation + text-encoder perf#18

Open
forkni wants to merge 13 commits into
SDTD_031_devfrom
feat/v2v-rcfg-correctness
Open

fix: realtime param/prompt correctness + cosine_weighted interpolation + text-encoder perf#18
forkni wants to merge 13 commits into
SDTD_031_devfrom
feat/v2v-rcfg-correctness

Conversation

@forkni

@forkni forkni commented Jun 13, 2026

Copy link
Copy Markdown
Collaborator

Changes

  • 30cc1db feat: add defensive OOM/size-mismatch fallback in pipeline call with scheduler rebuild
  • edb5834 fix: refresh _alpha_next/_beta_next on t_index runtime update; warn on do_add_noise=False bleed regime
  • 1d3a7c8 fix: init stock_noise from init_noise.clone() for coherent RCFG warm-start (matches reference fork)
  • 55de6c4 fix: preserve stock_noise across seed blends; sync _init_noise_rotated (eliminates per-frame RCFG cold-restart)
  • 74fefa2 chore: align V2V config defaults with reference fork (cache_interval=4, similar-filter on, do_add_noise true)

Branch

feat/v2v-rcfg-correctness -> SDTD_031_dev

forkni added 13 commits June 12, 2026 21:38
…n do_add_noise=False bleed regime

_update_timestep_calculations (stream_parameter_updater.py) now keeps the
pre-computed shifted tensors _alpha_next, _beta_next, _init_noise_rotated in
sync after a value-only t_index_list update.  Previously they were only built
in prepare() and the error-fallback _refresh_derived_tensors(), causing stale
RCFG-self stock_noise rotation at guidance > 1.0 when the user changed
Tindexblockstep at runtime (F2 fix).

Also adds a logged warning when do_add_noise=False is combined with
use_denoising_batch and a high-noise inter-step timestep (beta_sqrt > 0.75).
With do_add_noise=False the inter-step x_t_latent_buffer carries zero noise
content (alpha_sqrt * x0_pred only), causing the UNet to ghost-bleed previous
frames once beta_sqrt crosses ~0.78 (empirically at t_index_list[1] < 30 in a
50-step LCM schedule).  Root cause: do_add_noise=False violates the on-manifold
assumption; reference fork default is True.

Diagnosis confirmed via beta_curve.py (tmp_debug/, not committed): beta_sqrt
ranges from 0.61 at idx=36 to 0.78 at idx=29, matching the observed threshold.

6 CPU-only regression tests added (tests/unit/test_derived_tensor_sync.py).
@forkni forkni changed the title fix: V2V/RCFG correctness — stock_noise warm-start, config defaults, scheduler resilience fix: realtime param/prompt correctness + cosine_weighted interpolation + text-encoder perf Jun 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant