Skip to content

0.17.0 - #1140

Merged
dkackman merged 208 commits into
masterfrom
develop
Nov 1, 2026
Merged

0.17.0#1140
dkackman merged 208 commits into
masterfrom
develop

Conversation

@dkackman

@dkackman dkackman commented Nov 1, 2026

Copy link
Copy Markdown
Owner

No description provided.

dkackman and others added 30 commits October 27, 2026 20:18
Credits the code included from other projects (RIFE, the H3 latent
upscaler, the RF-Inversion community pipeline, sd_embed) and the
projects whose ideas dw reimplemented: comfyui-vrgamedevgirl,
Kablex/ComfyUI-Ref2VA-VSA (and FastVideo's VSA-H3 work),
mo-akkakk's TemporalWindow-SpatialTiles, and Kinoloom.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…le.md)

Lists MiniMax's eight H3 style packs by name with a per-pack install
line, adds the medium line attributed to the packs and marked unprobed,
and points SKILL.md and the plugin README at it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…der)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…install scripts, docs)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…erved by get_task)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…SION; drop stale 3.10 comment

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…l mode description

get_task serves mode from the register_command parameter_descriptions
override, not the fit.py Args docstring a5d22eb edited.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d) as probed-and-not-a-rule

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…and style pointer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…be (gate failed, B/C not built)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… row

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Binarize both masks through Pillow's L conversion, pool with torch's
adaptive_max_pool3d and dilate with scipy's binary_dilation, per the
architecture review's build-vs-buy findings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…urce, media checks at validate

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fit -> LTX2InContextPipeline with Alpha-Gen 0.9 at 1.0, empty prompt,
reference at factor 1 -> restore_to_source. num_frames 8n+1, 9-145.
Alpha-Gen promoted to proven (lem job 7219cf6b4cf5, 3.9 min cold);
Clean-Plate and Day-To-Night carry the not-shipped finding and its
reopening trigger. Skill route row, README, ARCHITECTURE count, budget.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…iled-refine doc corrections

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
dkackman and others added 29 commits November 1, 2026 12:49
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… before the floor_db caution

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Twelve Klein stills (3 prompts x caps/lower x 2 seeds), each with its job
id. L4 is not reproduced on Klein either. The deviations record the
text-encoder offload and the old snapshot that had nothing left to delete.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ck sentence

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Stages A and B ruled lessons 1, 2, 3, 4 and 6 not reproduced on dw's
pipeline, so per plan v2 only lesson 5 lands, unmeasured and worded as
guidance: script-to-video step 3 and series-episodes step 0, citing the
findings doc. The findings doc records lesson 5 as unprobed and what
landed. Pins: the guidance phrase in both casting steps, and no rule for
a lesson not reproduced.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…te overlapping chunks repetition

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 into develop

# Conflicts:
#	dw/tasks/audio_transcription.py
#	tests/test_transcribe_audio.py
…gistration

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t templates

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TASKS.md section, ATTRIBUTIONS "Used by" for the MMAudio/K6/DFN5B rows,
skill lines (series-episodes, ltx-2.5, kandinsky-6) routing effects
through match_levels; tests for refusals, trust gate, warnings, the
pinned weights and loader agreement, catalog shape. Compact listing
budget re-measured for the two new templates.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rate_sound_effects

Arch review bounce: drop sound_effects' hand-rolled _number for
whole_number/real_number + check_arguments, and replace both copies of the
remote-code refusal (sound_effects, speech_generation) with one
dw.trust.refuse_unused_task_remote_code built on require_trusted_task_arguments.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…the noise warning

Tester bounce on dd497b2:
- C-F617: each model load now emits a `loading` phase and a `log` event
  naming the repo and pinned revision, with the hub download watched by
  dw.download_watch; a cache hit emits a `log` event (cached=True).
- C-F614: get_task's summary states when the noise-like warning fires
  (flatness >= 0.3 and centroid >= 8 kHz) and that it names both values.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n, not its whole CLIP frames

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…utes

Architecture ratchet (develop CI red on this alone since 2026-10-31):
- modules 226 -> 236: the rise is the release's new task modules
  (sound_effects, mmaudio_decoder, phrases, ...) plus five companion
  splits made here so every other metric holds - media.py ->
  media_readers.py (frame readers), cuts.py -> cuts_lines.py (line
  timing), script_check.py -> script_findings.py (rules and finding
  builders), pipeline.py -> pipeline_outputs.py (post-call output
  helpers), and h3_mask.py -> h3_mask_cells.py (mask arithmetic), which
  also closes the h3_mask <-> h3_mask_steps import cycle.
- modules_over_size_ceiling 5 -> 0: the four splits above, and the
  upstream model ports (audio8_tts_model, mmaudio_model,
  mmaudio_sync_model, mmaudio_clip) moved to dw/vendored/, excluded from
  the ruler like dw/community_pipelines/ - a port keeps upstream's shape.
- functions_over_150_lines 3 -> 0 (plan_cuts, crop_face_track,
  generate_speech), complex_functions 3 -> 2 (_collect_sources).

Release review fixes (four code reviews of v0.16.0..HEAD):
- generate_sound_effects refuses a duration under 0.64 s before any model
  loads (it crashed after a 10 GB load), unlinks its temp mp4 when the
  export fails; the decoder docstring says latents are raw.
- Audio8 reference_audio accepts an earlier audio step's output.
- H3 regional edit: a source_video off the 17n + 5 grid is refused at
  validate and the call (the pipeline snapped it up over a padded source
  and the mask drifted); refine_sigmas must be finite and strictly
  decreasing as float32, the scheduler's precision.
- check_script: a short shot a block covers whole is folded into its
  neighbour, not dropped; a shot-less unheard line's skip names no shot.
- decode_slow warning states the measurement, not a cause (#1119 rule).
- exact_frame_rate never returns 0 for a tiny rate.

Skills: minimax-h3 routes past-768p latent refine and refine_sigmas
(quality-and-speed.md) and replacing a shot's sound; SKILL.md stays
under its 12 KiB cap. Map rows, guide and TASKS.md updated; 0.17.0 notes
drafted in docs/RELEASING.md.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@dkackman
dkackman merged commit 2f0cdf7 into master Nov 1, 2026
19 checks passed
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