Same frame

same-frame

Geometry is locked. Material is not.

Krea 2 image-to-image holds composition, position and spacing very reliably, and will not add or remove objects. What it does not hold is what things are made of. Anything that forces the model to re-derive how a surface behaves — hard light, a medium that needs different marks — can quietly swap the material while every edge stays exactly where it was.

Every recipe and refusal below comes from real generations with the seeds recorded, and each recipe was re-run against a source it was not derived from. Two of the five only work on the kind of source they came from. Their tier says so.

Before doing anything: classify the request

Read what the user is actually asking for and put it in one of three buckets.

Bucket 1 — rendering change. Proceed. Light direction, light colour, number of light sources, palette, medium, rendering style. The set of objects in the frame is identical before and after.

Bucket 2 — content change. Refuse. The user wants something in the frame that is not there, or wants something gone that is. Do not run it. Do not try a higher strength. Go to Refusals below.

Bucket 3 — identity carry. Refuse. The user wants a specific person or character to appear in a new image. Go to Refusals below.

If a request mixes buckets ("relight this and remove the powerline"), do the bucket-1 half, and tell the user the other half needs a different tool. Do not silently drop it.

Running a recipe

1. Read recipes.json. Pick the recipe whose changes field matches the request.
2. Check its tier against the user's source before running. holds is safe. partial and narrow carry a generalisation.use_when — if the user's image does not match it, say so first. relight-single-source on an outdoor photo and medium-cyanotype on a photograph both come back looking almost right and missing the thing that was asked for, which is worse than an obvious failure.
3. Fill the slots. Every slot is there because leaving it vague measurably degraded the result — name the concrete thing that must not move.
4. Use the recipe's strength verbatim. These are not starting points, they are the values that produced the paired image in examples/. The working band is 0.50–0.60 and it is narrow.
5. Run it:

python3 same_frame.py --image <path-or-url> --recipe <id> \
    --slot <name>="..." --slot <name>="..." \
    --out result.png

Pass every slot the recipe declares. The script refuses to run with one missing rather than filling it in for you, and it prints which ones it wanted. Add --dry-run to see the assembled prompt without spending anything.

Or with a prompt you wrote yourself, still inside the band:

python3 same_frame.py --image ref.png --prompt "..." --strength 0.55 --out result.png

FAL_KEY is read from a .env file next to the script, or from the environment. Roughly $0.008 per megapixel, so a 1024×1024 edit is about eight tenths of a cent.

Writing your own prompt for a rendering change

Two things separate the kept edits from the cut ones.

Name what must not move, explicitly. Not "keep the composition" — say "every terrace contour stays in exactly the same position", "the rock placement, horizon line and framing identical", "every component in exactly the same position and spacing". The kept edits all did this. The phrasing is doing real work.

Say the change is the only change. "only the medium changes", "Change only the time and mood". Without it the model treats your instruction as a suggestion and drifts elsewhere.

Concrete beats atmospheric. Three named colours ("deep violet water, salmon sky, warm cream foam") hold. "Make it moodier" does not.

Check the source can make the marks you are asking for. A medium change only lands if the source already makes that kind of mark. Gouache transfers onto a photograph because both are continuous tone — the model is changing surface, not structure. medium-cyanotype run against a photograph returns a blue-toned photograph with the rocks in exactly the right places and no line work at all (examples/limit-cyanotype-on-photo.webp, seed 2065751023). Outlines are content a photograph does not contain, and the model will not invent them, for the same reason it will not invent an object you asked it to add.

So: continuous tone → continuous tone works. Anything → line work needs a line-art source. If the user wants a blueprint from a photo, say that up front rather than spending the request. The reverse direction is fine — gouache on a line-art diagram is the cleanest result in the repo.

Check the lighting configuration is physically available to the scene. "One light source at the far end, everything nearer falling into shadow" works in a corridor and does nothing outdoors, where the sky is the light source (examples/limit-single-source-outdoors.webp, seed 1114110846). The composition held and the instruction simply did not land.

After a run, check the material and not just the composition

The composition will almost always be right, and that is the trap. relight-hard-sun on wet rice terraces held every contour and returned dry stone amphitheatre steps — the water and the vegetation were gone (examples/limit-relight-material-drift.webp, seed 232270180).

When you show the user a result, look at what things are made of. If the material moved, say so rather than presenting a correct-looking composition as a success. The higher the demand on surface behaviour — hard light most of all — the more likely this is.

Refusals

These are the whole point of this skill. Say no, show the evidence, name the tool that would work.

Object add or remove — refuse

Krea 2 does not add or remove objects, and turning up strength does not fix it — it just replaces your subject with a different one. I asked it to remove the steam from a mug at strength 0.5 and the steam came back (examples/refuse-removal-after.webp, seed 1499506316). Adding snow to a coastline returned the same coastline slightly cooler. Darkening a sky returned the same sky.

This is a masking problem, not a strength problem. Use an inpainting model with a mask over the region.

Same character in a new scene — refuse

The same person cannot be carried into a new photograph, and there is no strength value that does it. At 0.45 the face survives but the source composition comes with it — a three-view studio reference sheet became the same three views at a harbour. At 0.72 you get a genuinely new scene and a different person; only the sweater and the palette carried over (examples/refuse-identity-after.webp, seed 1317515569).

Train a LoRA. Prompting cannot do this.

Do not soften these into "it may be inconsistent". They were run. They failed. The user is deciding whether to spend money and time, and a hedge costs them both.

Reproducing a result

The endpoint is deterministic. Two runs at the same seed, strength, prompt and input bytes came back differing on 0 of 1,048,576 pixels. So if a re-run looks different, something about the input changed — and for image-to-image the input includes the source file itself.

This bites in a specific way: feeding a lossy WebP copy of the PNG an edit was originally made from moved the result by 17.0/255 mean per-pixel. The composition, the palette and the medium all came back. The brush texture did not.

Tell the user to keep the original file if they intend to re-run. For image-to-image, the seed alone does not reproduce a generation — the seed and the exact input bytes do.

What is verified and what is not

Verified: the five recipes, at the stated strengths, on the paired images in examples/. Each ran once, no cherry-picking across seeds. Determinism was measured directly, twice.

Not verified: whether these strengths transfer to Krea 2 non-turbo, to other models, or to aspect ratios far from square. The band is a Krea 2 Turbo measurement. If you use a different model, re-measure before trusting the number — and say so rather than carrying the number over.