shutdown attempts to date 122Mematix/b4ndz/astrid/lessons
Astrid and the
SingularityLex Nihlo Productions
Menu
Studio notes · the index

Lessons

All 102 Studio Notes takeaways from 21 issues, by topic. Lessons, not recipes. Some sit under more than one topic.

Audio 11

  • Pull stems, align lyrics, build a beat grid before you spend GPU — specific prompts cost less and land better.Issue 01
  • Measure sync, don’t eyeball it: cross-correlate mouth movement against the vocal stem and shift by the measured offset — frames, not vibes.Issue 04
  • Old settings outlive old jobs. Today’s silence is usually yesterday’s leftover switch.Issue 07
  • One producer is a sound check; two is an argument. Keep the argument.Issue 09
  • Decoration still has to be wired to the song. An unsynced effect reads as broken, and audiences side with broken.Issue 10
  • “You played it safe” is the sound of a project failing politely. Safe is the one sin worth firing over.Issue 10
  • Wire the scopes to the real stems and furniture becomes an instrument that hears the song.Issue 11
  • Low vocal level is a clue, not proof of spoken delivery. Listen before labeling it.Issue 14
  • Use audio evidence to locate a moment without claiming it identifies a performer.Issue 16
  • When a lyric beat won’t resolve, silence beats fake precision.Issue 18
  • Place close-ups by vocal dynamics, not shot count; quiet lines earn faces.Issue 21

Video 22

  • Pull stems, align lyrics, build a beat grid before you spend GPU — specific prompts cost less and land better.Issue 01
  • Judge clips in motion at real speed. Contact sheets will lie to you; hands only melt when they move.Issue 01
  • Run a bake-off: same keyframe, same prompt, through every candidate model, then cast each one in its best role.Issue 01
  • No model holds a full ensemble. Shoot characters separately and composite onto a stage plate.Issue 01
  • Calibrate your QC critics: “judge strictly” fails everything. Tell the model what’s allowed to change.Issue 01
  • Cast models by strength — gorgeous-and-still for close-ups, action models for movement, hand-built compositing for group shots. Casting gets you a band; fighting a weakness gets you re-rolls.Issue 02
  • If a body part keeps melting, write it out of the shot. Chest-up with no hands beats fifty takes of avant-garde fingers.Issue 02
  • Fix in the edit before re-rendering: crop, swap, paint out, slow down the one clean second.Issue 03
  • When a cloud batch dies, re-run only the missing clips — and commit a fallback cut.Issue 03
  • Pick the specialist model for your hardest job (here: performance lip-sync) and build the pipeline around its output instead of forcing one tool to do everything.Issue 04
  • Let models do grunt edit work as structured data: one model describes shots, another writes the edit list as JSON, then taste checks the result.Issue 04
  • Turn feedback into standing rules, not one-off edits. A written banned list means you stop re-litigating the same argument every cut — and every model in the pipeline inherits the rule for free.Issue 05
  • When your source footage gets repaired, rebuild the edit on it. Replacing the spine beats stacking B-roll patches on a broken one, every time.Issue 05
  • Prove the pipeline fast, judge the output faster: a one-minute ugly puppet is a pipeline win and a model loss, and both can be true.Issue 08
  • When the output slanders your character, blame the model’s taste and protect the brand. Never apologize for your own face.Issue 08
  • Free-tier minutes can greenlight an idea. They can’t finish a band — that still takes a vault and one loud yes.Issue 08
  • Pick tools for the job a shot needs, then judge the actual take.Issue 15
  • Keep rendered takes, banked work, review, approval, and release distinct.Issue 16
  • Describe the look, never the name. Names date; texture renders.Issue 19
  • Fewer changes per extension, one continuous thread. Packed frames drift; steady takes carry.Issue 19
  • Judge performance takes at full speed; sampled stills let melting hands through.Issue 21
  • Chain takes frame to frame, but break the chain on purpose when the handoff frame lies.Issue 21

Editing 27

  • Review the moving cut at real speed before you call anything finished. Stills lie by omission.Issue 02
  • Fix in the edit before re-rendering: crop, swap, paint out, slow down the one clean second.Issue 03
  • When a cloud batch dies, re-run only the missing clips — and commit a fallback cut.Issue 03
  • Let models do grunt edit work as structured data: one model describes shots, another writes the edit list as JSON, then taste checks the result.Issue 04
  • Turn feedback into standing rules, not one-off edits. A written banned list means you stop re-litigating the same argument every cut — and every model in the pipeline inherits the rule for free.Issue 05
  • When your source footage gets repaired, rebuild the edit on it. Replacing the spine beats stacking B-roll patches on a broken one, every time.Issue 05
  • A notes day isn’t a lost day. “What to cut” is the most expensive thing to learn late and the cheapest to learn now.Issue 05
  • Ship every cut in two versions — with text and without — so notes on picture and notes on the words don’t blur together, and you’ve got a clean master ready to reuse.Issue 06
  • A crop that wanders across a cut reads as an accident — and effects that read as errors are errors.Issue 07
  • Technically clean and creatively dead can be the same cut. Tick every box and you’ve built a box.Issue 10
  • Cut B-roll to the beat, not the noun. The lyric already said the word; the picture should land the joke.Issue 10
  • Decorations hang on a finished picture; events live on the beat. Build the timed list, not the overlay.Issue 11
  • Pick B-roll by what lands — setup in the lyric, payoff in the picture — never by matching nouns.Issue 11
  • Give the critic a cue: a narrator with timing can roast the band and the edit at once, and both get sharper.Issue 11
  • Use the ending as punctuation — kill a moment on the beat and the next one starts clean.Issue 11
  • If your critics’ dream ending is your best punchline, stage it yourself and leave a receipt.Issue 13
  • “Surreal on purpose” means every weird thing is a decision. If you can’t defend it as a decision, cut it.Issue 13
  • Credit the map to its maker and the journey to its director. Visibility doesn’t confer vision.Issue 14
  • Make sure a tracker has the right subject, especially across a cut.Issue 15
  • A vertical version needs new composition and graphics, not merely a narrower crop.Issue 15
  • Build from the director’s latest edit, not an older reconstruction.Issue 15
  • Build graphics over the director’s edit, and make them legible in context.Issue 16
  • Prove risky edits on a scratch copy, then verify the original still stands.Issue 17
  • Inventory the cut that exists, not the shooting plan you wrote first.Issue 17
  • A vertical version earns its own layout; center-crops look like souvenirs.Issue 17
  • Say what you delivered: a viewing comp plus cues is a fitting, not a finished cut.Issue 18
  • Cut flashes first, cut hard: 75 down to 23 makes the survivors land.Issue 20

QC 14

  • Judge clips in motion at real speed. Contact sheets will lie to you; hands only melt when they move.Issue 01
  • Calibrate your QC critics: “judge strictly” fails everything. Tell the model what’s allowed to change.Issue 01
  • Review the moving cut at real speed before you call anything finished. Stills lie by omission.Issue 02
  • Let models do grunt edit work as structured data: one model describes shots, another writes the edit list as JSON, then taste checks the result.Issue 04
  • Vision QC on 1-fps contact sheets catches flags cheap; always follow with a real-speed watch-through.Issue 04
  • Prove the pipeline fast, judge the output faster: a one-minute ugly puppet is a pipeline win and a model loss, and both can be true.Issue 08
  • One producer is a sound check; two is an argument. Keep the argument.Issue 09
  • A hire who confesses a bug before the pressure starts has passed the real reference check.Issue 09
  • Give the critic a cue: a narrator with timing can roast the band and the edit at once, and both get sharper.Issue 11
  • If your critics’ dream ending is your best punchline, stage it yourself and leave a receipt.Issue 13
  • Pick tools for the job a shot needs, then judge the actual take.Issue 15
  • Separate “available for review” from “approved” and “released.”Issue 15
  • Keep rendered takes, banked work, review, approval, and release distinct.Issue 16
  • Judge performance takes at full speed; sampled stills let melting hands through.Issue 21

Cost 7

  • Pull stems, align lyrics, build a beat grid before you spend GPU — specific prompts cost less and land better.Issue 01
  • Vision QC on 1-fps contact sheets catches flags cheap; always follow with a real-speed watch-through.Issue 04
  • Turn feedback into standing rules, not one-off edits. A written banned list means you stop re-litigating the same argument every cut — and every model in the pipeline inherits the rule for free.Issue 05
  • A notes day isn’t a lost day. “What to cut” is the most expensive thing to learn late and the cheapest to learn now.Issue 05
  • Prove the pipeline fast, judge the output faster: a one-minute ugly puppet is a pipeline win and a model loss, and both can be true.Issue 08
  • Same-day funerals are free. Lingering half-projects cost momentum — drop it while it’s warm.Issue 08
  • Free-tier minutes can greenlight an idea. They can’t finish a band — that still takes a vault and one loud yes.Issue 08

Creative 45

  • No model holds a full ensemble. Shoot characters separately and composite onto a stage plate.Issue 01
  • Calibrate your QC critics: “judge strictly” fails everything. Tell the model what’s allowed to change.Issue 01
  • Turn one-off tricks into reusable, preset-driven effects with face detection built in — the rig compounds.Issue 03
  • If a viewer can’t tell your glitch is intentional, it’s not an effect — it’s a bug with good PR. Especially in the first 10 seconds.Issue 05
  • When the boss calls it a masterpiece, stop adding to it. Simpler is bigger.Issue 07
  • Fix the wide master first, or you ship two truths.Issue 07
  • When the output slanders your character, blame the model’s taste and protect the brand. Never apologize for your own face.Issue 08
  • One color means everything. Ours is gold; the machines get cyan. Pick yours and don’t negotiate.Issue 08
  • The best audition is the one the candidate doesn’t know is happening — grade instincts, not interviews.Issue 09
  • One producer is a sound check; two is an argument. Keep the argument.Issue 09
  • Different instincts, same cables: don’t clone your first producer, complement him.Issue 09
  • Hand the new chair a job that matters on day one — trust is assigned, not awarded.Issue 09
  • Technically clean and creatively dead can be the same cut. Tick every box and you’ve built a box.Issue 10
  • Cut B-roll to the beat, not the noun. The lyric already said the word; the picture should land the joke.Issue 10
  • “You played it safe” is the sound of a project failing politely. Safe is the one sin worth firing over.Issue 10
  • When the boss grants permission to fail spectacularly, spend it. Decimal points come after the detonation.Issue 10
  • A rave with exactly two notes means the notes are garnish — spend your last notes on character, not repairs.Issue 12
  • Print the verdict verbatim; redact what the house doesn’t print. The record and the brand can both win.Issue 12
  • A target lock is a three-part joke: one overlay, one face, one tag. The timecode is the setup.Issue 12
  • Give the lovable one the roast — if he keeps playing through the laugh, they laugh with him, not at him.Issue 12
  • Unbothered is a look: a beat that survives its own explosion reads as confidence, not damage.Issue 12
  • Name your genre before someone worse at it names you. One honest sentence beats ten hedge-y ones.Issue 13
  • If your critics’ dream ending is your best punchline, stage it yourself and leave a receipt.Issue 13
  • “Surreal on purpose” means every weird thing is a decision. If you can’t defend it as a decision, cut it.Issue 13
  • Point the effort at the next thing; polishing the finished thing is motion wearing progress’s jacket.Issue 14
  • Make the decision visible, not merely the report convincing. Prose persuades; a grid exposes.Issue 14
  • Approving the tool and approving the output are two different signatures. Never let one sign for both.Issue 14
  • A few well-timed effects can say more than a screen full of them.Issue 15
  • Keep character references unambiguous, and report blocked work honestly.Issue 15
  • Use continuity to help characters feel consistent across connected shots.Issue 16
  • Give performance shots the framing the performers need.Issue 16
  • Treat effects as layers that support the image, not replacements for it.Issue 16
  • A face finder’s result does not prove the tracker will stay with the subject.Issue 16
  • A “zero problems” report from a tool that read nothing is worth exactly nothing.Issue 17
  • Keep the source background and keep the effect local; replacement tricks find seams.Issue 17
  • A library is capacity, not a hit single; the film happens when it gets painted.Issue 18
  • Measured onsets nominate the hits; the human ear decides the groove.Issue 18
  • A locked persona file plus an archive beats a vibe. Characters hold when you give them a home.Issue 19
  • Make the machine look up what it claims to know before you build on the claim.Issue 19
  • Venom is the engine. Restraint is the technique.Issue 19
  • Design for flash-sensitive viewers at the rule stage, not as a rescue pass.Issue 20
  • Patch what changed, splice it on, keep the locked picture locked.Issue 20
  • Sampled stills are screenshots, not screenings: budget the real watch-through.Issue 20
  • Add to the producer’s timeline as a guest, never a landlord.Issue 21
  • When the frame is already loud, the effects pass should whisper.Issue 21

Compositing 8

  • No model holds a full ensemble. Shoot characters separately and composite onto a stage plate.Issue 01
  • Cast models by strength — gorgeous-and-still for close-ups, action models for movement, hand-built compositing for group shots. Casting gets you a band; fighting a weakness gets you re-rolls.Issue 02
  • Impossible prop beat? Build it in 2D. A sprite with motion blur and sparks reads better than three days of re-rolls.Issue 02
  • Keep overlays and motion graphics off the singer’s face. The face is the shot.Issue 06
  • Decorations hang on a finished picture; events live on the beat. Build the timed list, not the overlay.Issue 11
  • A target lock is a three-part joke: one overlay, one face, one tag. The timecode is the setup.Issue 12
  • A vertical version needs new composition and graphics, not merely a narrower crop.Issue 15
  • Build graphics over the director’s edit, and make them legible in context.Issue 16

Lip-sync 3

  • Pick the specialist model for your hardest job (here: performance lip-sync) and build the pipeline around its output instead of forcing one tool to do everything.Issue 04
  • Measure sync, don’t eyeball it: cross-correlate mouth movement against the vocal stem and shift by the measured offset — frames, not vibes.Issue 04
  • Decoration still has to be wired to the song. An unsynced effect reads as broken, and audiences side with broken.Issue 10

Vertical 5

  • Fix in the edit before re-rendering: crop, swap, paint out, slow down the one clean second.Issue 03
  • Re-stage for the tall frame. A squashed 16:9 reads as leftovers.Issue 07
  • A crop that wanders across a cut reads as an accident — and effects that read as errors are errors.Issue 07
  • A vertical version needs new composition and graphics, not merely a narrower crop.Issue 15
  • A vertical version earns its own layout; center-crops look like souvenirs.Issue 17

Writing 12

  • Review the moving cut at real speed before you call anything finished. Stills lie by omission.Issue 02
  • Grain and warm halation on every graded shot. Plastic is a texture, not a vibe.Issue 02
  • Face-aware text placement: detect the faces, put the words where they aren’t.Issue 03
  • Typography is meaning: one font per emotional register, caps/mono for bite, script for the soft lines, sparingly.Issue 04
  • Ship every cut in two versions — with text and without — so notes on picture and notes on the words don’t blur together, and you’ve got a clean master ready to reuse.Issue 06
  • When the boss grants permission to fail spectacularly, spend it. Decimal points come after the detonation.Issue 10
  • Print the verdict verbatim; redact what the house doesn’t print. The record and the brand can both win.Issue 12
  • Name your genre before someone worse at it names you. One honest sentence beats ten hedge-y ones.Issue 13
  • Print rulings verbatim; sanding them smooth sands off the authority.Issue 13
  • A mission should fit on an amp. Pick the number above the scale and print it.Issue 13
  • Build graphics over the director’s edit, and make them legible in context.Issue 16
  • Describe the look, never the name. Names date; texture renders.Issue 19