Where did this picture come from?
A user makes 154 pictures for one shot. A week later someone else opens the one they kept and asks how it was made. Right now the only way to find out is to ask them, and if they are not around the next person spends a day retrying what the first already knows does not work.
I hit this on Flow Sessions and got buried in my own scroll. Even when I found the picture again I could not say why I made it that way. This tool answers that question without the maker in the room. Thirty seconds to see what was run and what was kept, then the next person carries on from that recipe instead of starting the shot over.
Glossary
Everything that goes into one press of the button. The words written, the pictures fed in, which model, which settings. Change any one of those and it is a different recipe.
A user presses go once. That is one attempt. The provider charges for that press and hands back whatever it hands back, sometimes one picture, sometimes sixteen. The 4-press try in this scene returned 16 pictures and cost $0.32.
It gets a permanent ID. When one press hands back four pictures, the ID is how to point at this exact one.
An input to the model: a style to copy, a character who has to stay the same, a colour palette. Order matters to some models.
For any picture: which words made it, which references went in, which model, who ran it, when. When the trail is broken the page says so instead of guessing.
A user sees a picture they like and carries on from there. The tool remembers where they started, so the path back never gets lost.
How far out the view goesFour zoom levels, and the contact sheet is one of them
Everything below is one shot. That is 154 pictures from 16 recipes, made in about half an hour. A scene has 15 to 30 shots. The contact sheet below is the widest view and it earns its place. The whole run reads at a glance, and the colour shifts alone show where the work turned. What the tool will not do is scatter those same pictures as loose boxes and arrows. At this count that stops being a map and becomes a ball of wool. The map that does work is on screen three, and it groups recipes by the ingredients that went into them.
Screen one: every shotWhich shots exist, and how each one is doing
One card per shot. The picture on the card is the one the user kept. If nothing has been kept yet, the card says so instead of showing a random pick.
no generations
Screen two: inside one shotWhat was tried, in what order, and what is still alive
This is the screen a user lives in. Time runs left to right. Each column is a burst of tries. Inside a column, recipes that only differ a little get stacked together, so 154 pictures read as about a dozen decisions instead of a wall.
Screen three: the mapWhich recipes share ingredients, and when each one ran
There is no picking what to compare by looking at thumbnails. Two pictures can look identical and share nothing. Two pictures can look completely different and come from almost the same recipe. So this map places recipes by what went into them. Each row is a set of recipes built on the same style picture. Position along the row is the actual time it ran.
50 pictures
8 pics
8 pics
8 pics
8 pics
8 pics
8 pics
2 pics
24 pictures
8 pics
8 pics
4 pics
4 pics
16 pictures
16 pics
16 pictures
16 pics
16 pictures
16 pics
32 pictures
16 pics
16 pics
A mode, not a screen: line them up by modelOnly earns its place when more than one model is in play
This one is niche, so it drops from a top-level screen to a mode that gets switched on. Each row is one recipe, each column one model, and a dash means that pairing never ran. On this shot it happens to be worth showing, because four models ran and three recipes crossed between them.
| recipe (and what went in) | midjourney-v7 | midjourney-v8 | krea-2-large | seedream-5-pro |
|---|---|---|---|---|
| 304b1513fc styl·c1473c same recipe, other model |
— |
|
|
— |
| 4c430985b1 styl·2d9f67 same recipe, other model |
— |
|
|
— |
| 315ca06ab3 styl·2d9f67 char·33a5bb same recipe, other model |
|
— | — |
|
| d046452a33 styl·c1473c pale·406bae char·33a5bb |
|
— | — | — |
| ce7811467e styl·c1473c pale·406bae |
— |
|
— | — |
| aa39be3923 styl·c1473c char·33a5bb |
|
— | — | — |
| 2759f3a1be styl·fe5d5f |
— |
|
— | — |
Screen four: one picture, everything about itShot Details, with one amendment
This is what opens on a click of Shot Details from screen one, and there is a wrinkle worth deciding now. A shot holds many pictures, so Shot Details has to pick one. It should pick the kept one, since that is the shot's answer, and it should carry the shot's totals alongside it: how many recipes, how many pictures, what it cost, who worked on it. That gives two different clicks on the same card. Clicking the picture opens the work on screen two. Clicking Shot Details opens the record. If nothing has been kept yet, it opens on the totals with the picture slot empty.
char_sheet 1ae6ae0d ─┼─▶ this attempt ─▶ 16 outputs
parent ─ not recorded
| this picture | 461038eca0c1362f |
| this try | 20260806T062412Z_midjourney-v7 |
| this recipe | 252f82b4acdea442 |
| model asked for | midjourney-v7 |
| model that ran | — not written down |
| who ran it | midjourney |
| their job number | — not written down |
| seed | None |
| times pressed | 4 |
| pictures back | 16 |
| cost | $0.32 |
the pictures fed in, in the order they went in
f1f082404f4e
1ae6ae0d4a47
Screen five: what changed between these twoThe one nothing else does
Two tries, words that changed on top, pictures underneath. The open question is how anyone gets here. A pair cannot be chosen by squinting at thumbnails, so the way in is the map on screen three. Find two recipes that share ingredients, then open them here. The map is the selector and this is the reading room.
How it actually gets usedFour jobs. The first one is how we judge whether any of this worked
If someone cannot do the first one, nothing else matters.
How the tool decides two recipes are relatedAnd why it will never put a similarity score on the words
The map groups by ingredients because ingredients are written down. Words are harder, and getting this wrong would quietly teach a user something false.
Measuring how much text changed is the wrong measurement
Take the two recipes from the compare screen. They differ in medium, colour grade, film stock, lens, lighting and style picture. Compared as whole blocks of text they score 0.893 similar, which reads as nearly identical. Scoped to the one section that actually moved, they score 0.355. The template swamps the signal.
One word can carry the whole shot
Swap black for asian and a different person is standing in the frame. Swap fast for slow motion and the shot reads differently end to end. Off-the-shelf text-similarity models score those swaps as almost no change at all, because they are built to treat that kind of substitution as a rephrasing. They are answering a question nobody here asked.
So the rule is: show the words that changed
A number like 0.94 sitting beside two recipes says they are basically the same. Sometimes they are the same, and sometimes one word turned the whole shot over. Reading black → asian settles which case it is immediately. The tool shows the change and leaves the judgement to the person looking at it.
The question actually worth asking
Small change, big difference in the picture. That pair is the most useful thing in the whole archive, because everything else stayed still by accident, which is as close to a fair test as real work ever gets. It is also the pair a similarity ranking would bury, because it looks like a near-duplicate.
How this feeds back into making thingsThe part that makes it worth building at all
This one is not urgent, so here is the thinking rather than a spec. Right now a shot is an elephant only its maker knows how to ride. They know how it moves. Anyone else climbs on and crashes into a wall. The point of the whole thing is that someone opens that work, finds a picture that landed, and starts from its recipe. That has three consequences and they all point the same way.
The recipe has to travel without the picture
A kept picture is a starting point, so its recipe has to come away clean and run again on its own. That is what the Remix button is. It also means the map should filter down to kept recipes only, which for someone arriving cold is probably the most useful view in the tool: here are the doors that open.
The ingredients have to still be there
A recipe that cannot run is a museum piece. Style pictures currently live at file paths and temporary web links, and temporary links expire. A shelf of starting points that fail when run would be worse than having no shelf, so keeping a permanent copy of every ingredient stops being tidiness and becomes the thing that makes branching possible.
New work lands next to what it came from
When someone branches from a recipe, their attempt inherits most of its ingredients, so on the map it appears right beside the thing it grew from. The map thickens around the work that succeeded. That is the tree, and it only draws itself once the tool writes down what each try branched from.
Labels the tool has to carryEach one changes what the tool is allowed to claim
These are not decoration. Get one wrong and the tool starts making promises it cannot keep.
| label | what it can say | why it has to exist |
|---|---|---|
| how much trail | all of it · some of it · none of it | All 16 real ones are some of it. A label is cheaper than a wrong answer. |
| can this model repeat itself | no · takes a seed · reports the seed · usually repeats · proven to repeat | A plain yes-or-no already gets Midjourney wrong. This decides whether the compare screen is allowed to offer a real test. |
| how the work carried on | ran it again · copied the recipe · used the picture · combined two · started fresh | Copying a recipe and feeding in the picture are different things. Lumping them together hides that an input changed. |
| why a strand stopped | dropped · the round ended · it has gone quiet | A strand with nothing after it might be rejected, unfinished, or just resting. Three different meanings. |
| what a try is doing | queued · running · finished · failed | Where it came from is written down before it runs, so a crash leaves a recoverable try instead of a picture with no history. |
Four things I am refusing to buildAnd why each one would do harm
These are the real decisions in this proposal. Everything else is layout.
1. No loose scatter of every single picture
I had this one too broad at first. A map is fine when distance means something recorded. Screen three is a map, and it works because the rows are ingredients written down in the receipts. What stays banned is the other thing: 154 individual pictures floating in one space, positioned by a physics simulation. That looks authoritative while being unreadable, which is the worst combination.
2. Nothing sits near anything else by accident
An existing 3-D graph drew five tidy clusters. They looked like taste. They were just which board the pictures came from, with nicer names. Two pictures sat together because of where they were saved, not because they looked alike. When a layout pushes things together, people read meaning into it.
3. No leaderboard of which reference works best
It would only count the ideas someone chose to keep pushing. People stop early on the ones that look weak, and change the words at the same time as the pictures. A ranking built on that is both wrong and convincing. The tool reports counts with their totals beside them and leaves ranking alone.
4. No similarity score on the words
The tool shows which words changed and stops there. A score would put a one-word swap and a full rewrite on the same scale, and those two behave completely differently. Section 09 has the numbers from real recipes.
The lookTwo options, both already yours
I used the cream one, because this screen would sit right next to the package tool already in daily use and it may as well match. The dark one is here as the alternative.
purple = current
green = done
these are the exact colours the
package tool already uses
The purple and green already mean the right things in the existing screens, so nothing new to learn.
gold marks the thing that matters
this is the idea-harvest page,
built for picking through
thousands of things
Honestly a fair claim, since it was built for exactly this kind of sifting. It just belongs to a different project.
Five open questionsEach one is answerable without reading anything above
Which look?
Cream page with black text, like the package tool already in daily use. Or dark grey with gold, like the idea-harvest page. I built the cream one. Either is a day of work, so the choice is taste, not effort.
What should be on screen at open?
Every shot as cards, leaving the choice of where to go. Or straight into the shot last worked on. I guessed the second one, because the work tends to sit inside a single shot for hours.
Should the tool guess when a new round starts, or be told?
Right now it guesses by looking for gaps of more than two minutes, and it admits it is guessing. The other way is a “new round” button pressed before starting. One extra click buys a fact.
The compare screen: early or late?
It is the only thing here nobody else has, which argues for building it first. It is also the one that can confidently show a pattern that is not real, which argues for building it last, once the record underneath is solid.
Lots of small pictures, or fewer big ones?
This page is packed on purpose, so a whole shot reads at once. The trade is that every picture is small. Six big pictures instead of sixty small ones is a different tool, and now is the time to decide.