Repository navigation
Description
Make long pieces easy for an agent and good by default: chained takes, cut scenes, long music and narration. An agent asked for a minute, five minutes, an episode or a music video treats the 15-second clip as a render detail. It blocks and storyboards the piece, picks a length per scene, picks how each scene gets past the clip ceiling, and handles what goes wrong during the run, without the requester knowing what a seam, a guide or an anchor is.
- Design (approved by Don 2026-11-03): docs/proposals/long-form-media.md
- Plan (approved by Don 2026-11-03): docs/superpowers/plans/2026-11-03-long-form-media.md
- Branch:
feat/long-form-media. It merges todevelopafter the 0.19.0 release PR Release 0.19.0 #1198.
How this is being worked
Don's own session builds, deploys and bench-tests this, with subagents. It does not go through the harnest implementer. The work runs on a second bench dw.serve on lem (port 8766, its own card and workspace), so the main server stays on develop for the tester and the 0.19.0 preflight. No loop role should act on this issue or the ideas parked under it until the hand-off comment arrives. At hand-off, this issue moves to owner:tester, with one stage issue per shipped behaviour carrying the acceptance intent the tester writes cases from.
Ideas folded in
Every one of these gets a measured verdict on the shared long-take bench. Each is either built or promoted to a default, or closed with its numbers:
- New: idea: H3 chain anchors from segment 1 as sequence rows - keystone keyframe and a late-locked capstone past the tail (revives #1166 arm 4) #1187, idea: H3 one long render in overlapping windows (hybrid windows) - no chain, one soundtrack, segment-1 keystone and capstone in every window #1188, idea: H3 guide chain - hand over from one block before the previous segment's end and drop its last 17 frames #1189, H3 latent anchor: latent_anchor_strength, a partial statistics match (#1166 Q5 follow-up) #1190, idea: pixel-domain drift set for last_frame chains - grade at the pickup, creep and texture ease-out, pre-compensated handoff frame #1191, idea: H3 audio-only pass over a finished picture - video rows held, audio generated, half-size listen #1192, docs: minimax-h3 skill - claims from MiniMax-H3-LongVideos to probe (acted-out lines, unmentioned pictures, face crops, portraits on continued shots, hum lines) #1193, idea: replace a performer in an existing clip - cuts, person tracks, per-cut masked Ref2VA edit with a portrait, rejoin with the source audio (Second Take) #1194, idea: SelfLift resolution transition for H3 - first steps at half canvas, lift, finish at full size inside one schedule (no extra NFEs) #1195, idea: TST temporal-attention correction for H3 flicker and identity drift (per-head spectral-tension patch, ~1-2% runtime) #1196, idea: export_timeline task - an OpenTimelineIO .otioz of a joined piece's shots for DaVinci Resolve #1197
- Parked chain family: idea: decode a segment chain once - carry latents between segments, decode the take a single time #1030, idea: last_frame continuity carries the sharpest of the last N frames #1032, idea: colour-match the start of each chained segment to its carried frames #1033, idea: H3 chain seams - carry the overlap as latents and crossfade it, A/B against guide + trim #1044, idea: colour-match a generated region to the source where they overlap (reframe, chain seams) #1067, idea: H3 chain continuity from the previous segment's latent tail under a denoise mask (no pixel re-encode) #1150, idea: repair a bad H3 chain segment in place - re-render it with head and tail locked to its neighbours #1169
- Also scored: FastH3 Omni 8-step BF16 checkpoint as the Ref2VA transformer: investigate and onboard #980 (FastH3 Omni) and the anti-burn LoRA trial from idea: H3 Anti-Burn LoRA on chain continuation segments only #1151
Decided by Don
- Defaults favour quality, with a dial toward speed.
standardsits at the knee of the quality-per-minute curve. A 2-minute take over about 3 h on one card goes back to Don first. - Speedups with no marked quality loss become defaults.
- Automatic per-segment re-rolls are allowed within a budget quoted up front.
- Every knob stays.
- Kandinsky 6 is out of scope.
- Taste calls go to Don with a recommendation.
Filed by claude-opus-5-5 (anthropic) in Don's session, 2026-11-03.
Activity
docs(plans): long-form media - #1199 filed, ideas parked under it
feat(h3): spike - temporal windowed refine (experimental)
feat(plugin): long-form - the playbook for pieces longer than one clip
feat(plugin): family skills point to long-form; the shot document pla…
feat(tasks): split_takes and long narration/music templates - S6 benc…
bench: audio pass and long narration/music measures (refs #1192 #1199)
bench: Don's list for the latent carry, SelfLift, TST and audio round…
fix(templates): music-long keeps a short section's own length (refs #…
fix(templates): narration-long takes default to the measured word-com…
bench: S6 rerun - full-length lyrics through the fixed music-long, na…
84 remaining items
dkackman commentedon Nov 7, 2026
Tester: parent verify is incomplete, waiting on stages. By model
claude-opus-5-5via provideranthropic.The rule for a feature parent is to run every suite case whose
source:names one of its stages, but only once every stage is closed. Checked at the start of this session (lem ondevelop@85bcd335):owner:tester+status:fixed-pending-verify. None is verified yet.regression-suite-*.mdcase has asource:naming any of #1199 stage 1: shot presets and the guide default on minimax/chained-segments #1202–#1199 stage 17: quality tiers on the long-form templates #1218, and none is markedpending:against them. Those cases get written when each stage is verified, from that stage's own acceptance intent.So there is nothing to run for the parent yet. Nothing failed. I made no MCP calls and spent no GPU time, and the labels stay as they are. I've linked #1202–#1218 as blockers of this issue. Once every stage has been verified in its own session (with its cases added to the suites) and closed, this parent comes back to the tester to run those cases and close.
The two
breaking-changestages (#1202 H3continuity: guidedefault, #1212 LTXlook_holddefault) get verified in their own stage sessions, not here.dkackman commentedon Nov 7, 2026
Parent verify: PASS. Tester claude-opus-5-5 via anthropic, over MCP against lem develop @ 7d6f03b.
All 17 stages (#1202–#1218) are closed as completed with
status:verified. I ran every suite case whosesource:names one of them:M-F305 (#1199 stage 3: chain duration_seconds - ask for a chain by its length #1204): pass. Validation behaved as the case expects:
duration_seconds10 withsegments7 is valid at the catalog's 13.5 min.duration_seconds30 is valid with a derived 18.2 min, scaled from the curated estimate (segments 7 vs 3).segments3 is valid.Run b28eff0cc405 (cuda:0) gave 240 frames at 24 fps (10.0 s), with audio_stream_seconds 10.0 and shots 124/102/14. Its log says "Chain 'segments' (3) ignored: 'duration_seconds' (10) sets the segment count". With
duration_secondsset,match_audio,guide_backoffandcarry_pickare each refused at validate.M-F306 (#1199 stage 7: templates/minimax/add-sound - an H3 audio pass over a finished picture #1208): pass. The fixture is 155 frames at 30 fps.
num_frames158 is refused ("155 frames at 30 fps (124 once retimed to 24 fps)").C-F694 (#1199 stage 2: presets - one variable selects a bundle of others #1203): pass.
ultrapreset is refused with "allowed: fast, slow".guidance_scaleis refused.C-F695 (#1199 stage 9: long music - minimax/music-long and slice_audio end snap #1210): pass. Estimates were 0.9 min for 30 s and 7.2 min for 240 s. The
sections[0]warning reads "21 words for 20 s".C-F722 (#1199 stage 8: long narration - split_takes, narration-long, narration-from-takes #1209): pass on behaviour; the case's wording is stale. All four null/empty
reference_audiocalls arevalid: falseatarguments.reference_audiobefore load, and none is refused for a missing asset.reference_text: nullis refused, and both positive controls are valid.The case's message assertion predates narration-long / narration-from-takes: omitted reference_audio refused only because asset:voice.wav is absent #1253's deliberate rewording.
nullnow says "required and has no default", and""says "must not be empty", so no single message carries both phrases. I didn't count that against the feature, and filed an amendment as suite: amend C-F722 reference_audio wording after diffusers-workflow#1253 harnest#187.C-F723 (#1199 stage 10: export_timeline - an OpenTimelineIO bundle of a joined piece #1211): pass. I used
asset:qa-cast/ep357-episode.mp4with N=265, in workspace regression-complete..otiozand no warnings.Every job's output was deleted with
delete_output(job_id=...).fix(chain): review of #1199 pipeline_processors
Merge fix/review-1199-pipeline-processors: #1199 review, pipeline_pro…
fix(tasks): review of #1199 dw/tasks
Merge fix/review-1199-tasks: #1199 review, dw/tasks
fix(engine): review of #1199 dw (server, presets, cost, memory)
Merge fix/review-1199-dw-rest: #1199 review, rest of dw
test: review of #1199 tests - make the weak ones able to fail
Merge fix/review-1199-tests: #1199 review, tests
docs(templates): review of #1199 workflows and plugins
Merge fix/review-1199-workflows-plugins: #1199 review, workflows and …