krea2-wildcards — Review Evidence

Human-review contact sheets and their source renders for the krea2-wildcards prompt library: a project that compiles researched visual vocabulary — art styles, linework and coloring, poses, lighting, camera composition, character design — into ComfyUI Dynamic Prompts wildcards for the Krea2 model family.

This dataset isn't the wildcard library itself. It's the photographic evidence a reviewer looked at when deciding, for each catalog entry, whether a given prompt fragment actually changes the rendered image the way it's supposed to. 18,535 images, 22.8 GB: 463 contact sheets (each tiling 5–10 generated cases with the catalog item ID and seed printed under every thumbnail) plus the 18,072 individual 1024×1024 renders each sheet was tiled from.

Why this exists

Most prompt libraries publish a wordlist and ask you to trust that it works. This one publishes the evaluation trail instead: every catalog axis (a lighting rig, a linework style, a pose) was rendered across many seeds, tiled into a contact sheet, and judged pass/hold/reject by a human looking at the actual pixels. That verdict lives in the GitHub repo's catalog; the sheet and renders behind it live here, so the verdict is never a black box — anyone can re-open exactly what was judged.

Contents

Phase	Runs	Sheets	What was under test
Phase 6 — calibration	13	70	Small anchor probes run before each larger evaluation
Phase 6 — single-axis	6	60	One catalog axis changing at a time, everything else fixed
Phase 6 — pairwise	2	20	Two axes changing together, checking for conflicts between them
Phase 6 — presets / random / benchmark	3	13	Full-scene preset contracts, random-utility sampling, the fixed benchmark case
Phase 7 — mass promotion	12	258	Full-catalog promotion runs per axis (character design, background, camera, lighting, linework/coloring, pose, preset) plus their calibration probes
Artist research	4	11	Artist-signature and art-style-family screening/retest
Anchored rig	2	31	The corrected subject: a 20-item probe and the 300-item three-seed revalidation that promoted linework_coloring 268/300

Structure

reports/<run_name>/review/
  prompt_matrix_<mode>_<NNN>.png   # the contact sheet image(s)
  manifest.json                    # sheet -> case mapping: which item/seed is in which cell
  crib.txt                         # (4 runs) reviewer's per-case reasoning notes
  review_parts/sNNN.json           # (v2 runs) per-sheet verdicts, one file per sheet
  saturation_screen.json           # (v2 runs) per-item saturation against its palette group
reports/<run_name>/runs/<style_id>_seed_<seed>/
  image_01.png                     # the original 1024x1024 render behind one sheet cell
  run.json                         # seed, resolved prompt, generation provenance
reports/runs/<style_id>_seed_<seed>/
  image_01.png                     # 15 standalone smoke renders, see below
  run.json

run_name matches the report directory in the GitHub repo's tests/reports/, so a sheet, the originals it was tiled from, and the scorecard/summary that judged it are always traceable to the same run and the same catalog commit.

The one exception is reports/runs/ with no <run_name> above it: 15 renders of five style packs at three seeds each, generated as smoke checks outside any evaluation run. They have no contact sheet and no verdict attached — they're here for completeness, not as evidence of anything.

The subject

Every image depicts the same fictional benchmark figure — not a real person, and not modeled on one. She exists purely so that an axis (lighting, coloring, a camera crop) can be judged with everything else held constant. Wardrobe across the corpus is studio and editorial fashion test clothing: blazers, blouses, coats, dresses. Because this corpus spans dozens of independent evaluation runs collected over time, we recommend a scan for scope before relying on any single image outside its own run's manifest — the description above reflects what the runs were designed to test, not a frame-by-frame audit of all 18,535 images.

Known limitation in the early history

Every rig prompt through 2026-08-01 opened on an unanchored "one adult woman" — no age, no ethnicity specified — which the Krea2 model resolves to a middle-aged Western face by default. The catalog's actual intended subject is a young Korean woman with an East Asian face. Two consequences worth knowing before trusting an older verdict in this dataset:

• Every verdict recorded against a sheet from before that date judged its axis on a subject nobody had actually specified.
• At least one specific finding — that the deep_jewel coloring value only renders saturated when paired with broad_blended_plane shading, recorded reviewing phase7_linework_coloring_v1 — did not reproduce under a corrected, subject-anchored rig. It looks like an artifact of the old rig's photorealism instruction fighting the illustration instruction, not a real property of deep_jewel.

A subject-anchor fix landed in the GitHub repo on 2026-08-01 (fix: anchor the benchmark subject and repair five dead body rewrites). Two runs here were generated after it and depict the intended subject: pilot_subject_anchor_v1 and phase7_linework_coloring_v2. Every other run predates it.

The deep_jewel claim above is now settled rather than merely suspected. phase7_linework_coloring_v2 covers all 300 items at three seeds, and the six shading medians fall between 0.1541 and 0.1708 — crisp_shape_shadow, the pairing recorded as collapsing to greyscale, measures highest. The actual driver is accent_light: on deep_jewel, controlled_rim averages 0.1751 against reflected_color's 0.1237. The original finding compared a controlled_rim render against reflected_color ones and attributed the gap to shading. Its per-run saturation_screen.json carries the measurements.

That revalidation also found 56 of its own sheet verdicts to be thumbnail misreads once reopened at 1024×1024 — which is the practical warning for anyone using this dataset. A contact sheet is a navigation aid. Judge nothing about saturation or fine linework from one; open the render in runs/<test_id>/image_01.png.

Provenance

The GitHub repo's tests/reports/**/run-state.json records every image's resolved_prompt_sha256, workflow_sha256, and image_sha256, and the prompt matrix behind any run is deterministically re-exportable from the catalog at the matching commit — so every image here is independently reproducible, not just hash-verifiable.

License

Released under CreativeML Open RAIL-M. This license was chosen over a plain permissive license (e.g. CC-BY-4.0) because every image is a generative-model output depicting a human figure: RAIL's use-based restrictions (no use to defame, impersonate, or misrepresent an individual; no unlawful output) exist specifically for this situation, where a plain attribution license would carry no such safeguard.

This license covers the redistribution rights we hold over this collection and selection of images. It does not substitute for the terms of service of the model used to generate them (Krea 2 / krea2TurboOfficialComfy_krea2TurboMxfp8) — those govern the underlying generated content independently, per this project's own LICENSE-NOTES.md, and should be checked separately before any commercial reuse.

Citation

If this is useful, please cite the companion repository:

Krea2 Wildcards — a source-grounded prompt library for Krea2 / ComfyUI Dynamic Prompts.
https://github.com/innofree/krea2-wildcards