Seeds and output counts

A seed identifies a random stream. It is useful for replay and nearby exploration; it is not a portable quality score.

Seed plans

• SeedPlan.scout(count) derives a documented set of independent seeds for a broad first pass.
• SeedPlan.locked(seed) preserves one seed for a controlled run. The Python API can repeat it, but identical requests can produce duplicate candidates; Studio therefore pairs Locked mode with one output.
• SeedPlan.explicit([...]) uses caller-provided seeds exactly.

Every candidate records the actual integer used. A backend must not share one mutable generator across a batch and then describe the batch as independently seeded.

Choose an output count

Count	Use it when	Trade-off
1	checking a recipe or device	fastest, no meaningful comparison
4	normal scouting and variation	default balance of breadth and cost
8	a direction is still ambiguous	broader but slower and harder to review

The Simple view always uses four. Advanced exposes 1, 4, and 8 so the contact sheet and review flow remain predictable.

A useful designer loop

1. Scout four independent seeds with Balanced structure.
2. Pick on concept, not on a technical badge alone.
3. Vary the selected candidate with a structure lock and subtle strength.
4. Change one recipe dimension at a time when comparing runs.
5. Export before starting a materially different direction.

Do not publish a “recommended seeds” list as if the numbers generalize across models, prompts, schedulers, or inputs. Publish the complete recipe and result lineage instead.

A real heritage-seed check

The AWS Korea 2023 Coding School notebook at main@4b3a220 recorded 6764547109648557242 as its active first-exercise seed. Its same code cell preserved 6854547109648557242 and 6634547109688557242 in comments, alongside 40 steps, CFG 7, and Canny thresholds 140/160. AIsketcher 0.2 reran all three numbers against the current pinned SDXL Canny Lite recipe. It used the already installed cache in offline mode, so the comparison downloaded zero bytes.

The notebook contains two prompt records that should not be conflated. Its opening lesson text proposes 알파벳 A, 글자, 추상화, 멋진 펜아트; the saved stdout for the active-seed execution records 특이한 풍경, 환상적인 그림, 복잡한 나라, 귀여운 케릭터. The provenance file preserves both as the documented brief and recorded runtime prompt, respectively.

Only the three numeric seeds were carried into the v2 compatibility run. The input, prompt, model stack, and generation recipe deliberately changed:

Dimension	2023 notebook evidence	AIsketcher 0.2 compatibility run
Source	present_image1.jpeg	reviewed metadata-free Pocket Kingdom source
Prompt	saved Korean runtime prompt above	expanded English graphic-design prompt
Base / ControlNet	DreamShaper / SD Canny, revisions unpinned	SDXL Base / SDXL Canny Small, revisions pinned
Scheduler	PNDM	UniPC
Steps / CFG	40 / 7	24 / 5.5
Canny low / high	140 / 160	90 / 224

This is a seed-compatibility study, not a claim that v2 reproduces the complete 2023 recipe or its pixels.

Verified 2023 execution evidence · active seed 6764547109648557242 · 40 steps · CFG 7 · Canny 140/160. This is the image embedded in that execution cell, not the separate showcase JPEG.

The legacy stack did not pin model revisions, so this evidence is preserved as a historical result rather than promised as a pixel-identical v2 replay.

2026 Studio hero interpretation

The English Studio documentation uses a separate, visually directed v2 study. Its selected first direction keeps the exact active seed 6764547109648557242; the model, prompt, and rendering recipe are deliberately modern. The English prompt is a human-written synthesis of the lesson brief and saved runtime prompt, not a claimed recovery of the old AWS translation.

Setting	Studio hero study
Preset	pinned sdxl-canny-lite@1
Outputs	4 explicit seeds; no random seed
Selected seed	6764547109648557242
Steps / CFG	36 / 7.0
Canny low / high	140 / 160
Control strength	0.55

Current pinned v2 run. The first direction is the HPO-selected fixed seed used in the English Studio screenshot; the other three values are explicit and recorded in the manifest. This is a modern interpretation, not a pixel-identical replay of the 2023 stack.

The complete manifest and HPO provenance preserve the prompt, negative prompt, model revisions, runtime, settings, seeds, and artifact hashes.

AIsketcher 0.2 separately reran the three notebook values with its current pinned recipe. The current images are retained as compatibility evidence, not as a visual recommendation:

Open the current v2 compatibility contact sheet

Left: exact active 2023 seed and canonical reference. Center and right: the two preserved comment values. Select the sheet to inspect it at full size.

Seed	Structure	Edge cleanliness	Distinctiveness	Review
6764547109648557242	0.271662	0.616346	0.194801	exact recorded active seed
6854547109648557242	0.322524	0.811936	0.198649	three technical badges; not an aesthetic selection
6634547109688557242	0.181476	0.373457	0.187767	third comparison direction

The center seed scored higher on all three automated diagnostics for this run, but those diagnostics do not decide what is beautiful. The exact active 2023 seed remains the canonical reference. The important result is that the old number was accepted exactly as a 63-bit seed and the new execution produced a fresh manifest with pinned model revisions and artifact hashes.

The legacy input JPEG was not copied because it contains camera and GPS EXIF. The current study uses the reviewed metadata-free source. The separate legacy showcase JPEG is not claimed as the recorded seed's output; the embedded cell image above is the verified execution evidence. See the bundle's provenance contract.