Super Slop Productions

How it actually gets made

Iteration

Everyone shows the one that worked. Here is the pile it came out of: the rejects, the notes written while watching them, and what it cost. Thirty-three tries for one frame that holds is not a failure rate, it is the method: the machine proposes, and somebody has to sit there and refuse the other thirty-two. Every number on this page was measured, not estimated, and every take it counts is still on the drive.

742Takes rendered
19Shots locked · one in 39
84hMeasured render time
13.3mAverage per take

01 · The tool · where takes live

Slopdash™

This is our in-house tool. We built it because nothing on the market kept what we needed kept: not the pretty take, the whole pile: queued, rendered, hearted, killed, with the electricity bill running underneath. It does the filing, assembles a rough cut on its own, and lets us go back to any version at any time.

Which means it makes room for the parts nobody puts in a pipeline diagram: the hesitation, the what did that other one look like again?, the let’s try three alts, just to see. None of that survives when going back costs a morning. It is the reason we can try more things than before.

None of this survives in a folder of files. Every take is archived with the exact prompt, seed, references, engine and cost that produced it, which is why a shot can be recreated identically months later, and why the numbers on this page exist at all.

Slopdash, RENDERS: the takes of one shot, hearted, validated or rejected

A shot is not a sentence you type once. It carries its structure, its references, its negatives and everything the takes before it taught us, and it gets rewritten the moment one comes back wrong. Held by hand, that bookkeeping is the slow part of the day: you re-read your own notes to remember why a rule exists. Held by a model that keeps the whole context, it is the fast part, which is why the meter below has nothing to do with GPUs.

Model budget · Max 20× plan Snapshot, mid-production
  • E2_beams 81
  • D_dodge 73
  • B_roar 60
  • E1_unmask 59
  • I_Title 51
  • A_rooftop 48
  • F_fall 46
  • C_launch 43
  • G_return 40
  • H_ending 33
  • J_logo 26
  • 0_intro 26
  • I2_pullout 18
  • B_crowd 18
  • B_street 16
  • B_taxi 14

The other meter. Rendering burns GPU hours; writing the shots, the prompts and the tools burns model budget, and the weekly ceiling is the one that actually stops the day.

02 · The cut · kept vs killed

Why this take

Rejected take: the beams come from the wrong place
Killed

E2_beams · take 43
“laser beams should come from his head”

Kept take: the beams come from the head
Kept

E2_beams · take 13
47 takes on this shot before one held.

Same shot, same prompt file, a different take. The line under each reject is the review note, word for word.

Rejected take: camera far too wide
Killed

A_rooftop · take 35
“camera way too far: top-down wide of the whole block, hero is a speck”

Kept take: the framing that worked
Kept

A_rooftop · take 28
81% of this shot’s takes went in the bin. The worst ratio on the film.

Take with frozen fires
Killed

F_fall · take 32
“fires near the cars at the bottom of frame are frozen, not animated”

Kept take of the fall
Kept

F_fall · take 18
Every flame has to move. That is usually what kills a take.

03 · The room · review notes

Review notes, unedited

A note means the take got close enough to argue with. The other 409 died in the second it took to look at them, and nobody writes a paragraph about a shot that is simply wrong. So these are not the worst takes of the film. They are the ones that almost worked, which is exactly why they are the only ones that taught us anything.

“laser beams should come from his head. Also, the wound should appear only as the lasers are doing damage, not before. When the arm is cut, which should happen a bit sooner, the monster should be in max pain, he should be losing his balance. Let’s animate the fire on the cars.”E2_beams · take 43
“he lands on the roof too late, the exhaustion beat has no time to settle; the sky turns to sunset far too visibly. The wreck and the continuity are better than v021 though.”G_return · take 30
“reference for the camera move and the barely visible sunrise; he lands on the roof earlier so the exhaustion beat works. Only flaw: the monster’s continuity.”G_return · take 21
“camera way too far: top-down wide of the whole block, hero is a speck. v028 framing is the approved one — close and low, roof visible, he is centred and fills frame.”A_rooftop · take 35
“he backs up walking, one foot then the other — both feet should slide together, pushed by the blast pressure.”D_dodge · take 54
“steering wheel missing, car was driving. Wanted: stopped, monster’s feet close, camera tracking him.”B_street · take 1
“cut 1: diagonal down-left to just above the core. Cut 2: short vertical up the left shoulder joint.”E2_beams · take 44
“charge-up validated: drains the arcs, then pumps up with orange energy. Bus position wrong here.”B_roar · take 42
“car must be stopped at the kerb, framed through the OPEN DOOR aimed at the centre of the road.”B_street · take 2
“cut in the middle — must be ONE continuous take, climb to landing to the kneeling beat.”G_return · take 31
“monster arms hang motionless; crowd not fleeing away from it; small fires frozen”C_launch · take 31
“lasers shooting from the arms, not the head”E2_beams · take 41

04 · The crew · working with the machine

The AI assistant

What a model actually unlocks here is not image-making. It is memory of the reasons. Every rule in a shot exists because a take failed once: the beam that ricocheted, the bus nobody asked for, the flare that ate the logo. Six weeks later the rule is still in the prompt and the reason is gone, so somebody breaks it again. A machine that holds the whole thread turns a correction into a decision that stays made, and that is the difference between a pile that repeats itself and a pile that gets somewhere.

The other half is that it argues back with receipts, and gets told off when it is wrong. Both are in here. On the left, building the site you are reading; on the right, making the film. The studio lines are the directors’ own words, kept as they were typed. Claude’s are summaries, including the ones where it lost.

The site

Type, colour, framing, and the machine getting told off for adding what nobody asked for.

ArilèsDon’t keep the image and an edit request — that fails often for lookdev.
Claude“Keep this image, apply that look” is an edit request: the model negotiates between preserving and transforming, and drops one of the two. Rewritten as complete prompts: whole scene from scratch, optical block spelled out, reference passed for colour only.
Every defect I was aiming at disappeared. Each one traded for a different one, which is its own lesson.
ArilèsGood. Now the faces have to survive it.
ArilèsI didn’t ask for a lens flare. Take it out.
ClaudeMine, added on top. You asked for four things: anamorphic, depth of field, chromatic aberration, barrel distortion. The flare was the one thing you never asked for, and the one thing that broke the shot. The bar across the logo and the sun patch eating the service cards were both it.
ArilèsCan we have the people looking at each other a bit more? Right now there is almost no eye contact.
ClaudeThe prompt said where people stood and never where they looked, so everyone got a middle distance and the room read as seven strangers. Rewritten with a gaze map: for each person, whose eyes they meet.
One deliberate exception: the one on his phone. Without him the group reads as a portrait instead of a conversation.
And a correction before you find it yourselves: I ran one batch at 4K and high quality in the same call. It is not the resolution that costs, it is the quality flag. 3,150 credits spent where 135 would have done, for the same definition.
ArilèsCheck the favourites.
ClaudeCaught me out. I had put take A on the site; you had hearted B. Realigned. Your hearts are the verdict, not my reading of them.
ArilèsExplain to me why I have been repeating twenty times that I don’t want the text formatted like this.
ClaudeBecause I was applying the rule by hand each time instead of writing it down. Every page carried its own copy of the type scale, so each new page started from a stale one and I “fixed” it again. A rule that lives in your messages is a rule I will break on the next page. It now lives in one file the whole site loads: paragraphs run the full measure, no exceptions. Eight width caps deleted with it.
ArilèsWhy is this on a black background? Explain to me why, when I ask you to put your art-director hat on, you miss things like that.
ClaudeBecause I looked at the elements and not at the page. The section had inherited a dark ground from an older draft, and I dressed the new blocks to match what was already there instead of checking the page against the charter. The charter is paper. It is on paper now.
ArilèsI don’t like the scrolling — what did you do?
ClaudeScroll-snap. I turned a preference about section rhythm into something your wheel had to obey. Out the same minute.

The film

Shot rules born from failed takes, tools that lied, and the nights the director overruled the machine.

GabrielOk, great — lasers shooting from the eyes, great. But… why the bouncing beam?
ClaudeMy fault. The attack was written as ONE FAST DIAGONAL DOWN-LEFT, and the model read the direction as the beam’s geometry, so it drew a ricochet bouncing off the armour. The rule since: directions bind to the impact point, never to the beam, and every beam dies where it lands.
ArilèsWhat is that brand-new bus in the middle of the street?
And again — no median strip.
ClaudeI rewrote the locks twice: first naming every forbidden object, then as physical assertions: “the middle of the road is just more asphalt.” Both cooks kept the bus.
GabrielYeah, no. Still a bus in the middle, a white van, and a median.
ClaudeThen we opened the references. Image7, the monster’s firing state, had the whole street in it: pristine bus across the lanes, white van, planted median. Six takes of defects were a faithful copy of one reference, and the prompt never stood a chance: an image outweighs every negative. One crop later, the street came out clean on the first try.
ArilèsYay — no more median, no more bus in the middle.
GabrielI saved, and it’s still sitting at v1.1.0 — aren’t you updating it?
ClaudeSo I automated it: the changelog bullets pick the level: NEW means minor, [major] means major, plain fixes bump the patch.
ArilèsJust saved, no updates in the minors — what’s up?
ClaudeWrong twice. The release tag only landed after a successful push, and this repo has no remote, so every push failed and no tag was ever written. And my auto-bump only fired if the current version’s tag existed. Frozen at 1.1.0 forever, by construction. The tag now lands on the commit, the history got backfilled by hand, and since then, saving restarts the server so it runs the version it just named.
GabrielThe monster is on the floor? No — he’s 20 feet tall! Looking down on the hero!
ClaudeI had fed the monster’s POV a street-level frame as its view reference, and the model matched the camera height, ignoring the note that told it not to. Same lesson, second verse, same night.
ArilèsHonestly — maybe just the city image, the kid, and the monster. Nothing more.
ClaudeThree references. The shot held.

05 · Shots that fought back

Shot by shot

Takes rendered per shot 742 total · 19 locked
  • D_dodge 61
  • E2_beams 52
  • I_Title 51
  • B_roar 47
  • F_fall 41
  • A_rooftop 39
  • E1_unmask 37
  • C_launch 34
  • G_return 33
  • H_ending 28
  • J_logo 23
  • 0_intro 19
  • I2_pullout 18
  • J0_glitch 9
  • B_street 4
  • B_toss 3

Bars are scaled to the heaviest shot. The count is every take rendered for it, kept or killed, read from the tool on 2026-08-19.

Where the takes went 742 rendered
  • 19Locked3%
  • 93Killed, annotated12%
  • 630Killed on sight85%

A note means the take got close enough to argue with. Most never did.

How long a take takes 382 timed
  • 78<1
  • 1591-5
  • 375-15
  • 3415-30
  • 6730-60
  • 760+

Minutes on the rig, per take

Median 3 min, mean 13 min, 84 h of machine time in all. The gap between those two numbers is the whole story: the slowest quarter of takes eats 80% of the rig.