{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "flux3-klein-in-the-coves-style",
 "title": "FLUX 3 Image Klein, ranked blind in a game's own painted style",
 "dek": "One person, ranking blind, put FLUX 3 Image Klein, a new text-to-image model, above FLUX.1 [dev], the model our game paints with, on 3 of six real subjects in the game's own oil-painted sketch style, one short of the bar we wrote down first: not a win.",
 "published": "2026-10-20",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0 for the words and tables; the pictures carry their models' own licences (FLUX 3 Image Klein's FLUX Kommunity License v1.0, non-commercial use; FLUX.2 [klein] 4B's Apache 2.0), each named in this site's licence ledger, and are shown as research receipts",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/flux3-klein-in-the-coves-style/",
  "md_url": "https://research.strata2signal.com/flux3-klein-in-the-coves-style/index.md",
  "md_sha": "a9577309c6d6c356e349ee88ce75f1d24ee7f9ed0dda31e81529f79f2a16829f",
  "html_sha": "eb5bd195a9e79377bb0fca41e56ae00444a6d15c0f03139412db8c280575f309",
  "anchors_sha": "96a2cc6a10f5c989f8c863e7a36e1bcb4be31478270cb22006eaa4e48b34b11c"
 },
 "short": {
  "paragraph": "On its release day, 2026-10-15, we ran FLUX 3 Image Klein in 16-bit (BF16) and in its maker's 8-bit integer (int8) build on one NVIDIA GeForce RTX 3090 (24 GB) power-limited to 300 W, at 768 × 768 pixels, beside FLUX.1 [dev], the painter of our game RealKeep, and FLUX.2 [klein] 4B. One judge, blind, ranked BF16 above FLUX.1 [dev] on 3 of 6 subjects and below it on 3 (one picture each); our rule needed 4, so the result is NO-WIN and the game keeps its painter. Median seconds per picture: BF16 5.87, int8 2.73, FLUX.1 [dev] 22.77, klein 4B 2.84. The page shows all 54 pictures FLUX 3 and klein 4B made; FLUX.1 [dev]'s stay held by a ruling of ours.",
  "from_dek": false,
  "counts": {
   "words": 2776,
   "minutes": 13,
   "tables": 10,
   "kit": false
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "what-the-model-is",
   "heading": "What the model is",
   "level": 2,
   "span": [
    1443,
    2644
   ],
   "chunks": [
    [
     1443,
     2644
    ]
   ],
   "chars": 1201,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-we-tested",
   "heading": "What we tested",
   "level": 2,
   "span": [
    2644,
    4046
   ],
   "chunks": [
    [
     2644,
     4046
    ]
   ],
   "chars": 1402,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-style-and-the-six-subjects",
   "heading": "The style, and the six subjects",
   "level": 2,
   "span": [
    4046,
    5930
   ],
   "chunks": [
    [
     4046,
     5930
    ]
   ],
   "chars": 1884,
   "digest": "The style is stylized oil-painted realism with loose quick brushwork. Six subjects were tested: the Reef-Runner, the Salt-Father, the sea-fog breakwater, the stone chapel, Harbormaster Haldane, and Wrenna Sault. Each prompt includes the style clause and a subject clause without negative prompts.",
   "digest_skipped": null
  },
  {
   "id": "the-rules",
   "heading": "The rules, written down first",
   "level": 2,
   "span": [
    5930,
    7130
   ],
   "chunks": [
    [
     5930,
     7130
    ]
   ],
   "chars": 1200,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-it-ran",
   "heading": "How it ran",
   "level": 2,
   "span": [
    7130,
    8046
   ],
   "chunks": [
    [
     7130,
     8046
    ]
   ],
   "chars": 916,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "seconds-per-picture",
   "heading": "Seconds per picture",
   "level": 2,
   "span": [
    8046,
    8865
   ],
   "chunks": [
    [
     8046,
     8865
    ]
   ],
   "chars": 819,
   "digest": "Median seconds for seeds 202 and 303 were measured. FLUX.1 [dev] took 22.77 seconds, klein 4B took 2.84 seconds, FLUX 3 BF16 took 5.87 seconds, and FLUX 3 int8 took 2.73 seconds. The int8 build took less than half the BF16 time.",
   "digest_skipped": null
  },
  {
   "id": "what-the-judge-ranked",
   "heading": "What the judge ranked",
   "level": 2,
   "span": [
    8865,
    10935
   ],
   "chunks": [
    [
     8865,
     10935
    ]
   ],
   "chars": 2070,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-pictures",
   "heading": "The pictures",
   "level": 2,
   "span": [
    10935,
    28838
   ],
   "chunks": [
    [
     10935,
     28838
    ]
   ],
   "chars": 17903,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "with-the-labels-showing",
   "heading": "With the labels showing",
   "level": 2,
   "span": [
    28838,
    30235
   ],
   "chunks": [
    [
     28838,
     30235
    ]
   ],
   "chars": 1397,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-this-changes",
   "heading": "What this changes for the cove",
   "level": 2,
   "span": [
    30235,
    30592
   ],
   "chunks": [
    [
     30235,
     30592
    ]
   ],
   "chars": 357,
   "digest": "Nothing changes for the cove because FLUX.1 [dev] remains the game's painter. The results for FLUX 3 and int8 are noted for a new, separately registered test if the team asks for one.",
   "digest_skipped": null
  },
  {
   "id": "what-this-means",
   "heading": "What this result does and does not mean",
   "level": 2,
   "span": [
    30592,
    31148
   ],
   "chunks": [
    [
     30592,
     31148
    ]
   ],
   "chars": 556,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    31148,
    32089
   ],
   "chunks": [
    [
     31148,
     32089
    ]
   ],
   "chars": 941,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "more-on-this-site",
   "heading": "More on this site",
   "level": 2,
   "span": [
    32089,
    32363
   ],
   "chunks": [
    [
     32089,
     32363
    ]
   ],
   "chars": 274,
   "digest": "The site provides information on who makes FLUX and its licences, as well as a licence ledger for every model run.",
   "digest_skipped": null
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    32363,
    32836
   ],
   "chunks": [
    [
     32363,
     32836
    ]
   ],
   "chars": 473,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-06cfc5af",
   "q": "how does flux 3 image klein function?",
   "a": "The model turns a sentence into a picture using 5.6 billion parameters and a separate text encoder. It generates images by cleaning noise out of a picture one step at a time.",
   "cites": [
    "what-the-model-is"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-6a2ba60f",
   "q": "what models were compared in the test?",
   "a": "The test compared FLUX.1 [dev] in an 8-bit floating-point build, FLUX.2 [klein] 4B in BF16, and two builds of FLUX 3 Image Klein: one in BF16 and one in int8.",
   "cites": [
    "what-we-tested"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-2e08ce78",
   "q": "what visual style was used for the prompts?",
   "a": "The prompts used a style described as stylized oil-painted realism with loose quick brushwork, in the style of Dishonored game art, without text, letters, signatures, or watermarks.",
   "cites": [
    "the-style-and-the-six-subjects"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-5b7e0d4c",
   "q": "what was the final result of the blind ranking?",
   "a": "The result was a NO-WIN because FLUX 3 Image Klein BF16 was ranked above FLUX.1 [dev] on only 3 of the 6 subjects, failing to meet the requirement of 4.",
   "cites": [
    "what-the-judge-ranked"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-eebbfaf0",
   "q": "what impact did the test have on the game?",
   "a": "The test changed nothing for the game. FLUX.1 [dev] remains the game's painter, and the testing line is closed.",
   "cites": [
    "what-this-changes"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "qwen-image-turbo-in-the-coves-style",
   "why": "shares ground with § Seconds per picture · § Seconds per picture, by each engine's own clock"
  },
  {
   "slug": "fifty-seven-milliseconds",
   "why": "shares ground with § Seconds per picture · § Finding one: a step is a fixed-price good"
  },
  {
   "slug": "a-rig-your-friend-already-owns",
   "why": "shares ground with § More on this site · § The kit"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThanks to Black Forest Labs and the ComfyUI team for same-day support. Software: **ComfyUI** v0.41.0 (GPL-3.0), **PyTorch** 2.13.0 (BSD-style), **CUDA** (NVIDIA, proprietary). A small human team set the rules, made the blind ranks and signed the numbers; AI agents ran the harness and drafted this page.\n\nCorrections and later measurements will be added below, each dated (UTC), saying in plain words what it counts.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 12,
  "digests_kept": 4,
  "chips_written": 8,
  "chips_kept": 5,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 1,
   "figure": 4,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 9
  },
  "new_noun_tokens_checked": [
   "1",
   "12",
   "13",
   "148",
   "16",
   "16-bit",
   "2.73",
   "2.84",
   "202",
   "2026-10-15",
   "21",
   "22.77",
   "3",
   "300",
   "303",
   "306",
   "3090",
   "4",
   "44",
   "4B",
   "5",
   "5.6",
   "5.87",
   "54",
   "6",
   "72",
   "768",
   "8-bit",
   "91",
   "BF16",
   "BFL's",
   "C",
   "COMPLETE",
   "ComfyUI",
   "Dishonored",
   "Elo",
   "FLUX",
   "FLUX.1",
   "FLUX.2",
   "GeForce",
   "Haldane",
   "Harbormaster",
   "Image",
   "Klein",
   "Kommunity",
   "License",
   "MiB",
   "NO-WIN",
   "NVIDIA",
   "RTX",
   "RealKeep",
   "Reef-Runner",
   "Salt-Father",
   "Sault",
   "US",
   "W",
   "WIN",
   "Wrenna",
   "int8",
   "seed-101",
   "v1.0"
  ],
  "new_noun_tokens_withheld": 1,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "The test asks if a judge, looking blind at one picture per subject, ranks FLUX 3 Image Klein above FLUX.1 [dev] on at least 4 of 6 subjects in the style of RealKeep. Four model builds were tested in C",
    "detail": "figures absent: ['768']"
   },
   {
    "kind": "digest",
    "reason": "figure",
    "item": "Rules were registered before the first picture. A valid picture requires the registered model, steps, size, and prompt. A judge ranks four seed-101 pictures per subject from 1 to 4. A WIN occurs if FL",
    "detail": "figures absent: ['3']"
   },
   {
    "kind": "bullet",
    "reason": "new_noun",
    "item": null,
    "item_chars": 3,
    "withheld": "a string naming something the article does not",
    "detail": ""
   },
   {
    "kind": "chip",
    "reason": "figure",
    "item": "A win occurred if FLUX 3 Image Klein (column C) was ranked above FLUX.1 [dev] (column A) on at least 4 of the 6 subjects, provided at least 4 subjects had valid pictures from both.",
    "detail": "figures absent from cited spans: ['3']"
   },
   {
    "kind": "chip",
    "reason": "figure",
    "item": "The total run took 13 minutes 44 seconds. For the median speeds of seeds 202 and 303, FLUX.1 [dev] took 22.77 seconds, FLUX.2 [klein] 4B took 2.84 seconds, FLUX 3 BF16 took 5.87 seconds, and FLUX 3 in",
    "detail": "figures absent from cited spans: ['13', '44']"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 43,
    "text_sha": "149ee92eabba",
    "figure": "2.73",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Bare figure: 2.73 carries no unit and no 'a picture'. The page gives 2.73 seconds a picture as the median over seeds 202 and 303 only, because seed 101 also counts reading the prompt, for FLUX 3 Image Klein's int8 build.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Bare figure: 2.73 carries no unit and no 'a picture'. The page gives 2.73 seconds a picture as the median over seeds 202 and 303 only, because seed 101 also counts reading the prompt, for FLUX 3 Image Klein's int8 build. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 37,
    "text_sha": "ccf54ed64423",
    "figure": "300",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Does not name the card in full and reads as a fact about every 3090: the page says its one NVIDIA GeForce RTX 3090 was power-limited to 300 W, and the limit read 300 W before every column.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Does not name the card in full and reads as a fact about every 3090: the page says its one NVIDIA GeForce RTX 3090 was power-limited to 300 W, and the limit read 300 W before every column. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 297,
    "text_sha": "b00205a39bab",
    "section": "what-the-model-is",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Says the model 'uses a distilled version', where the page says we ran the distilled version (the base version was not tested). It also widens the licence: the page says people and companies under US$5 million a year in revenue, affiliates included, may also use its pictures commercially, subject to its terms, not 'commercial use' at large.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Says the model 'uses a distilled version', where the page says we ran the distilled version (the base version was not tested). It also widens the licence: the page says people and companies under US$5 million a year in revenue, affiliates included, may also use its pictures commercially, subj…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 223,
    "text_sha": "f4e7a371342d",
    "section": "how-it-ran",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "The memory peaks name no model: 'for BF16' on a page where FLUX.2 [klein] 4B also ran in BF16, while the page's sentence is about both FLUX 3 builds. They also lose the setting they were read under: 768 × 768, with ComfyUI's older loader (dynamic VRAM off).",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: The memory peaks name no model: 'for BF16' on a page where FLUX.2 [klein] 4B also ran in BF16, while the page's sentence is about both FLUX 3 builds. They also lose the setting they were read under: 768 × 768, with ComfyUI's older loader (dynamic VRAM off). Not a gate drop, and recorded here …"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 248,
    "text_sha": "097f269d3d04",
    "section": "what-the-judge-ranked",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Drops two conditions the page states. The top-two line sits under 'Reported, not registered (none of it is a claim)', and the digest leaves out its other half, the bottom two on both portraits. 'Pretty much indistinguishable' was the judge's answer after the reveal, on the labelled pictures, not a blind read.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Drops two conditions the page states. The top-two line sits under 'Reported, not registered (none of it is a claim)', and the digest leaves out its other half, the bottom two on both portraits. 'Pretty much indistinguishable' was the judge's answer after the reveal, on the labelled pictures, …"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 237,
    "text_sha": "bd4d844958ea",
    "section": "the-pictures",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Gives the corner-mark count as a plain count. The page says 10 of the 54 pictures carry a scrawl or a stamp like a signature, and 'We counted them by eye on the pictures shown here.'",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Gives the corner-mark count as a plain count. The page says 10 of the 54 pictures carry a scrawl or a stamp like a signature, and 'We counted them by eye on the pictures shown here.'. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 247,
    "text_sha": "32783ca7263e",
    "section": "with-the-labels-showing",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Overstates the hands: 'on certain subjects', where the page finds the fused or missing hand on one subject, Wrenna Sault, and says Harbormaster Haldane's hands hold together in both builds at all three seeds. It also gives the page's own hand descriptions as the judge's notes, and the Salt-Father line loses 'what I feel like this foe would look like'.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Overstates the hands: 'on certain subjects', where the page finds the fused or missing hand on one subject, Wrenna Sault, and says Harbormaster Haldane's hands hold together in both builds at all three seeds. It also gives the page's own hand descriptions as the judge's notes, and the Salt-Fa…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 192,
    "text_sha": "3ef86ef15edd",
    "section": "what-this-means",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-18",
    "why": "Garbled, and drops its scope: the page says the result does not cover the game's own 16-bit FLUX.1 [dev] or test BFL's 91 Elo claim, and is one judge's blind look at six subjects, one style, one picture each, on one card. It also states 'under non-commercial use' as fact where the page says 'which we read as non-commercial use', for FLUX.1 [dev] and FLUX 3 Image Klein only.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-19: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-18 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-18 (UTC), after the writer ran and after the gates had counted: Garbled, and drops its scope: the page says the result does not cover the game's own 16-bit FLUX.1 [dev] or test BFL's 91 Elo claim, and is one judge's blind look at six subjects, one style, one picture each, on one card. It also states 'under non-commercial use' as fact where the page says '…"
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 227,
    "text_sha": "35b5c2b63c55",
    "chip_id": "c-75b55160",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-19",
    "why": "Gives the corner-mark count as a plain count and widens it: the page says 10 of the 54 pictures carry a scrawl or a stamp like a signature in a bottom corner, counted by eye on the pictures shown here, and names no watermark.",
    "detail": "removed from this pack by hand on 2026-10-19 (UTC), after the writer ran and after the gates had counted: Gives the corner-mark count as a plain count and widens it: the page says 10 of the 54 pictures carry a scrawl or a stamp like a signature in a bottom corner, counted by eye on the pictures shown here, and names no watermark. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-10-19T02:17:14Z"
}