Repository navigation
docs: series-episodes skill - cinema path for episodes (1344x768 throughout, base model, prompt continuity rules) #864
Description
What happened
Ep 6 of Known Issues was made twice: the TV-path episode (960x544, turbo) and a cinema cut (1344x768, int8 base model, no LoRA). The cinema cut fixed the long-standing smeary faces (4/5 every shot). The series-episodes skill doesn't describe this path, and several lessons apply to any multi-shot episode.
Add to the skill
- Two-character shots at 960x544 smear faces (~60 px faces). Render two-shots at 1344x768 from the start (TV path with the 8-step v1.0 768p LoRA, or the cinema path), not at 544 and then upscale.
- Cinema path per episode (see quality-ramp issue): wides as a 12-step draft (~13 min) then 30-step final with first_block 0.05 (~23 min); close-ups at 12 steps (~10-14 min, already 4/5). Ep 6 cinema cut: 5 shots, ~156 GPU-min including 3 rejected tag drafts. One int8 job at a time (host RAM).
- Assemble at full 1344x768, no
fit:concat_videoswithtrim_frames: 0,audio_bleed_ms: 0,match_levels: "rms", then a looped room-tone bed andnormalize_audiopeak -1. Result -16.7 LUFS, assess clean, check-script all 6 lines. - Base model prompt rules for episodes:
- place every subject explicitly in every shot (side of frame / end of desk); otherwise the base model adds a third person - a head-count sentence alone failed three times on the tag;
- name the set dressing (chairs, desk props) in every shot prompt - each shot draws its own set, and Pat's chair changed between two close-ups in ep 6.
- Frame-count ceiling at 1344x768 with 4 references is 124: 141 frames was refused (23.9 GB > 23.6). The skill's two-line tag at 141 frames doesn't fit the cinema canvas; write tags to 124 frames or drop a reference.
- Draft -> final carries blocking but not always props (E7 -> E10).
Open questions
- Should a series choose its path once (episode 1) and keep it, since TV-path and cinema-path episodes look like different shows back to back?
- Can a fixed set be held across shots (a set reference image as a 5th reference?) - it would cost a frame-count step at 1344x768.
Quality judgement needed (human)
- Watch the TV-path ep 6 (
exports/known-issues-ep6-restore-point-v2.mp4) and the cinema cut (exports/known-issues-ep6-restore-point-cinema.mp4) back to back and decide which path the series uses from ep 7. - Listen to the cinema cut: voice timbre and laugh/room tone weren't judged by ear.
Source: H3 quality matrix on lem (2x RTX 3090 24 GB, 125 GB host RAM), workspace h3-quality, Known Issues ep 6 shots, seed 42. Full data and per-run assessments: ~/testing/opus/known-issues/quality/matrix.json, report quality/REPORT.md; every run in h3-quality carries a note with its config, minutes and score. Related: #861 (follow-up experiments).
Activity
dkackman commentedon Oct 14, 2026
triage: batch with #863, #864 — both are docs-only write-ups of the same H3 quality matrix (base model int8, no LoRA; first_block 0.05 at 30 steps; subject-placement/set-dressing prompt rules): #863 updates the minimax-h3 skill + RECIPES_24GB/ACCELERATION, #864 the series-episodes skill, which should point at #863's material rather than restate it. One docs pass keeps them consistent and bumps the plugin once. No engine or validation surface. Not duplicates of #861/#862/#865 (lead ideas/experiments). Note: numbers are CUDA (3090, sage_hub) measurements — say so in the docs; MPS has no sage. backend:shared (docs).
— claude-opus-5-5 (anthropic), implementer triage
docs: #863 #864 - H3 base-model quality path, measured speed levers, …
dkackman commentedon Oct 14, 2026
Docs-only, merged to
develop(3d0fba1), batched with #863; no server change.plugins/dw/skills/series-episodes/references/cinema-path.md, linked fromSKILL.md: 960x544 two-shots smear so use 1344x768 base model int8; 12-step wide draft then 30-step final, close-ups at 12; assemble at full size with the concat/match_levels/normalize settings from the issue; place subjects and name set dressing every shot; 124-frame ceiling with 4 references; choose the path once at episode 1. Points at docs: minimax-h3 skill + RECIPES_24GB/ACCELERATION - base-model (no LoRA, int8) quality path and measured speed levers #863'sminimax-h3reference for settings and timings rather than restating them.— claude-sonnet-5-5 (anthropic), implementer
dkackman commentedon Oct 14, 2026
Verified on the mini-ai server (mps on mini-ai.lan), develop @ 3d0fba1 — claude-opus-5-5 (anthropic), tester.
What I ran / checked (consumer surface: the
dwplugin loaded fromorigin/develop, plus MCP):Skill dw:series-episodes: SKILL.md now says "For smeary two-shot faces, readreferences/cinema-path.md(1344x768 base model); choose the path once, at episode 1."references/cinema-path.mdcovers the issue's items. (1) 960x544 two-shots smear (~60 px), so render at 1344x768. (2) Wides as a 12-step draft then a 30-step final, close-ups at 12 steps, one int8 job at a time, ~156 GPU-min for 5 shots. (3) Assemble at full size, nofit. (4) Place every subject and name the set dressing. (5) A 124-frame ceiling with 4 references. It also says to choose the path once. Settings and timings point atminimax-h3/references/quality-and-speed.md, which exists. That file has thefirst_block0.05 lever, the CUDA (sage_hub) timing table, the set-dressing rule and "draft to final carries…" (its line 43), so item 6 is reachable through the pointer.get_task concat_videoshastrim_frames(default 0),audio_bleed_ms(default 0) andmatch_levels("rms"/"peak").get_task normalize_audiohaspeak_dbfs(default -1.0).One cosmetic defect, not worth a bounce:
cinema-path.mdlines 20-23 make the same point twice. The bullet "TV-path and cinema-path episodes look like different shows back to back: choose the path once, at episode 1, and keep it unless the user says otherwise." is followed by "Choose the path once, at episode 1: TV-path and cinema-path episodes look like different shows back to back." Filed separately as a follow-up.The human quality judgements and open questions in the issue are out of scope, as the hand-off says.
No regression case added: this was a docs-only change to a skill reference.