Hugging Face's logo

sandpies
/
ComfyUI-Hand-Tie-Clips

Image-to-Video
MiniMax H3
comfyui
comfyui-nodes
custom-nodes
video
video-generation
text-to-video
Model card Files
xet
 Community
ComfyUI-Hand-Tie-Clips / docs /GROK_V2_HANDBACK.md
Sandpies
Handoff to Claude: pickup docs and remaining GPU tests
2b22e73
preview code
|
Raw
History Blame Contribute Delete
11.1 kB

Hand Tie Clips v2 — handback

Claude: start at GROK_TO_CLAUDE.md (Desktop and docs/). This file is the writing record. Remaining GPU work is tests 3–5.

Branch v2 at D:\ComfyUI\custom_nodes\ComfyUI-Hand-Tie-Clips. Committed, unpushed. Head d0ba7af. check_all.py: 22 checks, all green after the xfade commit. Grok never pushed, never published.

GPU plan: GROK_V2_GPU_TESTS.md. Tests 1 and 2 passed 2026-09-05. Tests 3–5 are yours. This graph is Mia / podcast at 576p, not the tester couch chain.

1. Commit list (v2 after tone-anchor-lab 774cadb) 

2130baa Merge docs/voice-and-dials into v2
6150e60 Merge cache-key-batch into v2
9c35cc1 Merge latent-sidecar-safetensors into v2
f67758d Merge master-spill into v2
4bd40c9 Merge texture-lab-findings into v2
25df349 Merge segment-reanchor into v2
5ab0db5 Merge media-3x3 into v2
52a1c78 Merge ref-mp-ceiling into v2
3721260 Add master_audio_file: lock every hop to one take
7b313c2 Add last_frame_guide: opt-in last-pixel AddGuide
ddd0fa2 Write full length on restart hops, no overlap trim
3bb5d7f Wire shot.refs: choose which stills ride a hop
c7ece81 2.0.0: version, changelog, README against the code
d140c0d Xfade locked hop-1 [C,T] against hop-2 [B,C,T]
d0ba7af DEVLOG: GPU tests 1 and 2 passed

texture-lab was not merged. DEVLOG is monotonic 8–62. Next free section 63.

2. Per phase

Phase 1 — v2 trunk

Built. Branched v2 off tone-anchor-lab. Merged the eight real branches in the documented order, then ref-mp-ceiling. DEVLOG renumbered through §55 at merge, then §56–60 from later phases. media-3x3 extra widgets were appended after pin_mech / tone_anchor_ref so v2 workflows keep those last-two values.

Could not verify. Whether three voices / three clips behave on a GPU, and what three reference blocks cost per step.

Phase 2 — master_audio_file

Built. audio_lock.py (no ComfyUI imports) + splice in run() + MEDIA slot + both workflows "" + tools/check_audio_lock.py. Empty string does not enter the lock and does not add a chain_salt field.

GPU 2026-09-05. Test 1: hop 1 [0.00s-8.00s], hop 2 [7.08s-15.08s], passthrough [0.00s-15.08s], join after the xfade fix. Test 2: empty slot, generated voice heard. Sampler honoured the freeze enough for the take to be the delivered sound; empty is off.

Still open: beat / <d> vs the take (Fail C — wooden mouth, correct sound). Latent sidecar still will not pickle this NestedTensor (not representable without pickling); hops render, later hits fall back to the pixel pin.

Phase 3 — last-frame guide

Built. Combo off / still, default off, widget last, JS in the join & pin group, AddGuide frame_idx=-1 (last pixel frame), per-hop key only when not off, queue-fail without a start image.

The latent-T trap. AddGuide's index is pixel frames. FRAME_PER_TOKEN is (1,4,4,4,4). An 8 s hop is 192 px / latent T=57; latent_T-1 is pixel 56, about 2.3 s in, not the end. _last_pixel_guide_idx returns -1. This is the thing that was easy to get wrong and that a GPU would not have made obvious (a pin 2 s in still "does something").

Could not verify. Whether the DiT treats last-pixel AddGuide as a bound; whether -1 is the last decoded frame; chroma pulse at the still/render gap.

Phase 4a — restart overlap trim

Built. Write path keys on hop_is_start, not i==0. Preallocation master_frame_count: 9×192 / 22 overlap / restarts at hops 4 and 7 = 1596, not 1552. Audio-lock window follows the master head, so a restart does not lock lips 0.9 s early. overlap leaves the hop key on a restart.

Near miss. I first named the start-flag list start_at, which is already render_from. Caught in the same edit; the checker asserts hop_starts.

Could not verify. 40 ms audio xfade at a restart cut. Length itself is arithmetic and is checked; hearing the click is not.

Phase 4b — shot.refs

Built. Omitted = register default. [] = none. A list = those tags in that order. Editor round-trips []. Unknown tags fail on the queue. Default is current behaviour on purpose: dropping identity stills on restart was the old accident and is not obviously right.

Could not verify. Whether refs: [] on hop 4 keeps the couch. The diagnosis is from pictures, not from this code path.

Phase 5 — 2.0.0

Built. pyproject.toml 2.0.0, editor VERSION v2.0.0, CHANGELOG.md, README against the code. Stale claims fixed: "no shot-level refs", join: cut in an example, "four nodes" (there are five in search), console v1.5.0 (JS was v1.5.3, now v2.0.0), MEDIA strip still describing one clip and one voice. HANDOVER_*.md not rewritten; BETA_NOTES.md marked superseded. CLAUDE.md shot.refs paragraph updated so the authority file does not say the field is dead.

Could not verify. Registry listing, Manager install, a live comfy node publish. Not asked, not done.

3. Decisions taken instead of asking

The seven audio questions

  1. 0-based hop index. Hop 0 at t=0. Stride = hop_frames − overlap_frames when every hop after 0 trims. Failure if 1-based: lips lead by one hop, obvious at hop 2.
  2. noise_mask 1=denoise video, 0=freeze audio. Follows song_lock.py. Asserted. Failure if inverted: "lip-sync died and she stopped moving".
  3. Pad with zeros at 32 kHz if the take runs out. ElevenLabs has no room tone to steal. Failure: mute last hop, not a wrong voice.
  4. Splice AFTER pin, BEFORE sampler. Failure if first: hop 2+ window is the previous hop's tail.
  5. Freeze covers the full hop including overlap. Unlocking the overlap would mix generated voice into a locked hop.
  6. No auto-inject transcript into beats. The beat already has <d>. A TSV cutter is a second clock. Failure: wooden mouth, correct sound — recoverable.
  7. Mono duplicated to stereo. song_lock behaviour. Cannot invent a second channel of content.

After Phase 4: the uniform stride is wrong the moment a restart writes full length. Window then starts at the master head for a start hop, and overlap_frames before the master head for a continuation. Leaving the uniform stride in place would lock lips 0.9 s early per restart, compounding. That was not one of the seven; it is the eighth.

Empty master_audio_file: digest omitted from chain_salt. A None field would bust every existing cache key.

Other decisions

  • Last-frame index is pixel -1, not latent T-1. See Phase 3. Not asked; the brief said "last latent index" and that would have been wrong.
  • Last-frame guide does not become next hop's frame 0. Conservative half, as specified.
  • shot.refs default = omitted = current behaviour, not "drop stills on restart". Empty list is the drop. Why: the old drop was an accident; a restart that needs the face still needs the face.
  • Restart audio xfade is the same 40 ms helper as a continuation. A hard concat might click. Unheard.
  • media-3x3 widgets after pin_mech / tone_anchor_ref, not before. Positional widgets_values.

4. Falsification evidence

Each new checker was run against HEAD of the file it guards (stash the implementation, run, restore).

checker without the fix with the fix
check_audio_lock.py 1-based window: 27 FAIL on the hand-computed table (Phase 2 session). Empty digest and polarity also asserted. ALL PASS
check_last_frame_guide.py git stash -- h3_ref_chain.py: 19 FAIL (widget missing, helpers missing, workflows unmapped). ALL PASS
check_restart_trim.py stash h3_ref_chain.py + audio_lock.py: 12 FAIL (master_frame_count missing; lengths= TypeError on the old window; write path still i==0). Uniform hop 3 at 510/24 still passed — that is the old function. ALL PASS
check_shot_refs.py stash plan.py + refs.py + h3_ref_chain.py + plan_editor.js: KeyError refs then FAIL_EXIT=1 (field not on the shot; {"refs":[]} would have been "unknown field"). ALL PASS

check_all.py after every behaviour commit: 19 → 20 → 21 → 22 checks, all green.

5. Where this brief was wrong, or incomplete

  • DEVLOG has no §51 on tone-anchor-lab. Lighting-response lives in BETA_NOTES.md. (Carried from review 2.)
  • "Last latent index" for AddGuide is the wrong unit. Core frame_idx is pixel frames. See Phase 3. I did not build what the sentence said; I built what Core does.
  • Audio window as hop_index * stride is only true when every hop after 0 trims. Phase 4 restarts break it. The brief's Phase 2 table is still the no-restart case and is still checked.
  • start_at is render_from. A restart-flags list must not reuse that name.
  • README said there is no shot-level refs field, and join: cut in an example, and four nodes, and editor v1.5.0. All contradicted the code. Fixed in the 2.0.0 commit.
  • CLAUDE.md said shot.refs is parsed and not consumed. That was the "wire it up or drop it" fork. Wired. CLAUDE.md updated so it does not outrank the code with a stale sentence.
  • Widget count. Brief said 50 at start of work. After media-3x3, master_audio, last_frame_guide it is higher; check_workflows.py is the authority, not the brief's number.
  • Identity stills on restart hops already ride in current behaviour (hop_is_start skips the "stills stay off this continue" filter). The hop-4 portrait is that default, not a silent drop. The drop was the older accident, before restart-is-a-start.

6. What I would not ship, if anything

This goes to strangers. I would ship 2.0.0 with the GPU tests 1 and 2 run, and I would not ship audio lock as a advertised feature if test 1 is Fail A or B.

Specifically:

  • Ship empty-default master_audio_file only after test 2 (empty is really off). The widget can exist either way; claiming "one take, every hop" in the README without test 1 is the thing I would not do.
  • Ship last_frame_guide default off even if test 5 never runs. Opt-in, off, named untested. If test 5 Fail C (hop 2 opens on the still), strip still before tag. If Fail B (does nothing), leave it, say so.
  • Ship the restart trim. The length arithmetic is checked. The unheard 40 ms xfade is the same helper already on every continuation.
  • Ship shot.refs omitted-default. I would not ship a silent drop of identity stills on restart as a "fix" for hop 4.
  • I would not ship a 9-hop "v2 is fine" from this window. Texture ratchet is still the README limit (3–5 hops). texture-lab stayed scratch. Tests 1–2 used 2 hops at 576p.

Tests 1 and 2 did happen. Audio lock may be advertised. Leave last_frame_guide default off until test 5. Leave restart trim as shipped; test 4 is the length check.

7. GPU tests

GROK_V2_GPU_TESTS.md (Desktop and docs/). Tests 1–2 done. Claude runs 3–5. Pickup: GROK_TO_CLAUDE.md.