realkeep diffusion research

the licence test is stricter than "did the checkpoint look fine": every model in a render's graph is checked, LoRAs included. that is not a formality — the pixel mediums in our own v2 mass-render record a perfectly clean checkpoint and load a LoRA whose licence forbids commercial use of any kind. a checkpoint-only filter would have published them. 661 published + 338 held + 262 withheld = 1261 considered.

the painters our worlds run — and what their licences say

first, the fact everything below stands on: realkeep ships no model weights — not these, not any. every painter is chosen and downloaded by you, on your machine, under that model's own licence. the three below are the ones our own worlds run and this page's renders came from.

the withheld table shows the models that failed our licence read. this is the other half, and the half you actually need: the painters whose work is published on this page, each licence read first-hand (2026-08), so the rigor points both ways.

the full readings, clause by clause, live in the licence ledger — each model's licence is the authority on itself.

albedobase xl, and what it does with a style

the house spread — one painter, seven ways of seeing24G

compares: drawing style

albedobaseXL_v21
42 renders · 6 subjects · 2026-07-10

the step ladder — how many steps does a foe actually need?24G

compares: step count

albedobaseXL_v21
42 renders · 6 subjects · 2026-07-10

loot icons — three icon idioms against one control24G

compares: icon idiom

albedobaseXL_v21
12 renders · 4 subjects · 2026-07-10

limner sketches — painterly against comic24G

compares: sketch medium

albedobaseXL_v21
7 renders · 6 subjects · 2026-07-11

the step bench — two sdxl painters, three step counts96G

compares: step count (8 / 32 / 64) × 2 checkpoints

albedobaseXL_v21, sd_xl_base_1.0
6 renders · 2 subjects · date not recorded

nine worlds, several painters, and every medium we could ask for

before and after — art that remembers what happened

compares: before vs after the world changed

albedobaseXL_v21
44 renders · 11 subjects · 2026-07-26

the full matrix — every world, every medium24G

compares: art medium

Juggernaut-XL_v9_RunDiffusionPhoto_v2, albedobaseXL_v21
167 renders · 35 subjects · 2026-07-11

the mediums grid — does the medium survive the painter?24G

compares: painter × world

Juggernaut-XL_v9_RunDiffusionPhoto_v2, albedobaseXL_v21
38 renders · 8 subjects · 2026-07-11

nine worlds, three painters24G

compares: checkpoint

Juggernaut-XL_v9_RunDiffusionPhoto_v2, albedobaseXL_v21
18 renders · 9 subjects · 2026-07-11

rerolls — the same world, asked twice24G

compares: painter × seed

Juggernaut-XL_v9_RunDiffusionPhoto_v2, albedobaseXL_v21
16 renders · 6 subjects · 2026-07-11

its foes, and the words that summon them

hint robustness — the iterations that drifted24G

compares: hint variant

albedobaseXL_v21
30 renders · 6 subjects · 2026-07-13

who is the sketch actually of?24G

compares: subject assembly (old vs ADR-0154)

albedobaseXL_v21
30 renders · 5 subjects · 2026-07-12

the five bands — a foe getting visibly worse24G

compares: the foe's health band

albedobaseXL_v21
18 renders · 3 subjects · 2026-07-13

foe hints — three phrasings against five foes24G

compares: hint phrasing

albedobaseXL_v21
15 renders · 5 subjects · 2026-07-13

rewriting the band cues — and the seed policy underneath24G

compares: cue wording + seed policy

albedobaseXL_v21
12 renders · 2 subjects · 2026-07-13

one foe, three candidate hints24G

compares: hint candidate

albedobaseXL_v21
8 renders · 2 subjects · 2026-07-13

the limner deck — sketching a fight as it happens

compares: limner prompt assembly

albedobaseXL_v21
42 renders · 7 subjects · 2026-07-26

the limner deck, v2 — consequences and negative prompts

compares: consequence clauses + negatives

albedobaseXL_v21
42 renders · 7 subjects · 2026-07-26

what the hardware and the clock actually cost

two rigs, the same six paintings24G vs 96G

compares: which machine painted it

albedobaseXL_v21
60 renders · 6 subjects · 2026-07-16

the cold-start tax — what an idle gpu costs the first player24G

compares: render order (cold vs warm)

albedobaseXL_v21
10 renders · 3 subjects · 2026-07-14

the hi-res refine, with and without the chat model resident24G

compares: chat model resident or not

albedobaseXL_v21
2 renders · 1 subjects · 2026-07-11

the campaign archive census — all 2828 of them

the research pages show a curated spread, but the archive is bigger than any page should be. so here is the whole thing counted: every render the live world made, grouped by house style, painter, step count and canvas. the 40 groups shown cover 2,824 of the 2,828; the tail is 4 renders across 2 smaller groups, all in the JSON. the full census ships as campaign-census.json.

two notes so the numbers stay honest. this census is a different archive from the honest total at the top of the page: the census counts what the live world painted in play; the bench counts the experiment renders. no render is in both. and where a held painter appears in these groups, it is counts only — no flux-dev image is published here or reachable from this page; the JSON ships the tallies, not the art.

11 experiments, 338 renders, not published

flux.1 [dev]'s licence grants its outputs in one clause and restricts them in another. we are not going to read that in our own favour on a page whose entire claim is that we read licences properly — so every experiment below, each of which contains at least one flux-dev output, is held out of the published site until counsel reads it. the pages and their images stay in the repository, unaltered and uncounted as published. if the reading comes back clean they are published; if it does not, they never are. this is the same withheld-with-reasons rule the table above runs on, applied to ourselves.