comfy-director

Comfy Director

Story first, pictures second. Multi-shot AI films fail in the edit, not the render: beats that don't cause each other read as a montage wearing a story's clothes, no matter how good each clip looks.

REQUIRED BACKGROUND: the comfy skill (CLI mechanics + the recipe path — apply/capture/foreach — for reusable multi-stage graphs). On any failed job, comfy-debug.

Order of operations

1. Concept — one idea a stranger can retell.
2. Screenplay — before touching the pipeline. Deliver it to the user first.
3. Shot list — each scene mapped to a continuity strategy.
4. Audio design — against the screenplay's emotional beats, not after picture.
5. Produce → conform → QC.

Skipping 2 and jumping to renders is the #1 cause of killed cuts.

Screenplay rules

• One protagonist the camera can follow. Recognizable in every shot.
• Causality test: every cut answers "and BECAUSE of that…" — never

"and then, somewhere else…". If a beat can be reordered without breaking anything, it isn't a story yet.

• Stranger test: someone with zero context must retell the plot in two

sentences ("X wanted A, B stopped them, they did C, so D").

• Beats carry timings that sum to the target runtime.

Concept rules

• The brand/product is felt, not name-checked: argue its philosophy;

in-jokes and community lore quoted on screen read as fan-service.

• A technique the client mentions ("style transfer", a model name) is a

spark, not a mandate — use it only if the idea needs it. Building the film around a technique is the tell of a vendor, not a director.

• Weirdness must compound: surprising subject AND object AND setting.

One weird element in a normal frame is decoration.

Prompting shots

• Photoreal shots: live-action cinematographer / film-stock language ONLY.

One illustrator/anime/comics name in a photoreal prompt poisons it. Illustrated references belong only in deliberately stylized shots.

• Motion prompts describe HOW the scene moves, not what's in it (see comfy

skill video gotchas — t2v/i2v model traps, SaveVideo, fps wiring).

Continuity toolkit

Pick per shot, in rough order of strength:

If a shot can't hold continuity, rewrite the screenplay so it doesn't need that shot — don't ship the break.

Audio

• Score = short designed cues joined at story pivots; one long generation

will not follow a dynamic arc (it inverts exactly where you need the peak).

• Place VO by math: stem duration vs beat window, before mixing.
• Duck score under VO; loudness-normalize the final mix (~-14 LUFS).
• Dialogue must lip-sync by construction. Drive the mouth from the exact

audio that plays (KlingAvatar sound_file) and cut between angles of the same lip-synced head — B-roll over a continuous VO reads as a montage, and i2v-invented mouths drift. "No lip-sync / doesn't feel real" is the classic rejection. (Generate audio in-graph — cloud LoadAudio can't see uploads; see comfy audio gotchas.)

• Lock voice & casting by EAR before the expensive renders. Voice/accent is

subjective and unreadable from a waveform — render a few cheap candidate TTS takes of one line, have the user pick, THEN bake the winner into the costly, transient_auth-prone avatar/lip-sync shots. Re-rendering all of them on a rejected voice is the avoidable burn. Same logic for the protagonist's look: approve the hero still before generating every shot from it.

• Verify the mix by measurement, not faith. You can't hear the render —

ffmpeg volumedetect each section: a "dead air" ending or a buried line shows up as a window near -91 dB. Confirm every beat is present at level before declaring it done.

Production ops

• Append EVERY submit — shots, keyframes, audio stems, resubmits — to

job_ids.txt (<job_name> <prompt_id>) at submit time. Append-only: a resubmit gets a new line (shot_s3_resubmit <id>), never an overwrite, so dead prompt_ids keep their shot identity.

• Keep LOG.md at the project root: an append-only event log, written **at

event time, not session end** — one line per submit/resubmit (+ failure code), per QC verdict (+ re-roll reason), and per decision that invalidates earlier work: recasts, prompt changes, which reference image or stem is now canonical. Frame-by-frame findings live in qc/QC_LOG.md; LOG.md indexes them.

  17:36 ERROR shot_s3 transient_auth -> RESUBMIT 27dd45b4 (same workflow, no re-login)
  17:44 QC shot_s2 FAIL background-crawl -> re-roll (same kf, new seed); recut beat if 2nd fails
  17:50 DECISION recast protagonist -> ref_ray.png canonical; vo2, vo6 stale, re-gen

• Recovery contract: a successor agent with zero memory must resume from

disk alone: LOG.md (what happened, why) → job_ids.txt (which jobs) → comfy jobs status (their states). A decision not in LOG.md dies with the session — completed-but-stale jobs (an old casting's VO stems) otherwise look identical to good ones.

• ffprobe every downloaded clip; normalize off-spec duration/dimensions

before conform.

• QC one frame from every clip against the screenplay before cutting.
• Exact runtime: trim/retime per-beat in ffmpeg, don't hope.

Common mistakes (each killed a real cut)