#1199 stage 11: ltx2/chained-segments holds the look by default #1212

Activity

1. dkackman

   added

   owner:testerTester's turn to actTester's turn to act

   status:fixed-pending-verifyImplementer fixed, awaiting tester verificationImplementer fixed, awaiting tester verification

   breaking-changeInterface change the tester must account for before re-testingInterface change the tester must account for before re-testing

   stageA child issue: one stage of a featureA child issue: one stage of a feature

   target:lem

   on Nov 7, 2026
2. dkackman

   marked this as blocking on Nov 7, 2026

   • Long-form media: one playbook, measured levers, quality-first defaults for pieces longer than one clip #1199
3. dkackman

   added

   target:mini-aiFound against a non-lem server (DW_TARGET=mini-ai); Don triagesFound against a non-lem server (DW_TARGET=mini-ai); Don triages

   and removed

   target:lem

   on Nov 7, 2026
4. dkackman commented on Nov 7, 2026

   

   dkackman

   on Nov 7, 2026

   OwnerAuthor

   More actions

   Verified on the mini-ai server (mps on mini-ai.lan, develop @ cac31945) by claude-opus-5-5 via anthropic.

   Note on acceptance cases: the hand-off says "run every pending: #1212 case", but no suite file has one, and git log -S 1212 shows none was ever written. I tested against the issue's own acceptance intent instead.

   What I ran:

   • get_workflow templates/ltx2/chained-segments variables_only=true: the variables include look_hold: true.
   • validate_workflow + run_workflow templates/ltx2/chained-segments {seed:1212} (job f26ff33457c4) succeeded. get_job_workflow is realized: variables.look_hold: true. The chain block reads continuity: last_frame, segments: variable:segments (3), trim_frames 2, crossfade_ms 80, look_hold: variable:look_hold, with num_frames 121. The description documents the default and says "pass false for the plain chain". It also states the 15-20 s cap.
   • run_workflow ... {seed:1212, look_hold:false} (job 7b07954b1934) succeeded with the same shots layout (121/119/119 frames, 359 total).
   • compare_outputs with the false run as baseline, at 2.0 s / 5.2 s / 10.2 s:
     • frame 48 (segment 1) is identical.
     • frames 125 and 245 (segments 2 and 3) differ: PSNR min 35.4 dB, SSIM min 0.944, mean abs diff up to 2.9. By eye there is no visible difference. So the default really applies look_hold to the later segments only, and false turns it off.
   • Wall time: per-pass 21.9 s (default) vs 21.8 s (false). Denoise took 528 s vs 521 s and decode 40.8 s vs 40.5 s. The default run was cold (load 30.8 s), and the false run was warm on the cached model. So the cost is about the same.
   • Adjacent case: validate_workflow with look_hold: "maybe" is refused at arguments.look_hold ("Cannot interpret 'maybe' as true/false").

   All three acceptance points hold. Closing as verified on mps; CUDA has not seen this.

   No suite case added: each arm is about 10 minutes on mps, and the stage was supposed to come with pending: cases that never got written. A spec gap for the lead to note if wanted.
5. dkackman

   added

   status:verifiedTester confirmed the fix via a real MCP callTester confirmed the fix via a real MCP call

   verified-on:mps

   and removed

   status:fixed-pending-verifyImplementer fixed, awaiting tester verificationImplementer fixed, awaiting tester verification

   on Nov 7, 2026
6. dkackman

   closed this as completedon Nov 7, 2026