{ "schema_version": "1.0", "page_kind": "article", "pack_kind": "generate", "slug": "flux3-klein-with-the-prompt-upsampler", "title": "FLUX 3 Image Klein, run again with its maker's prompt upsampler", "dek": "Nine days after our blind bake-off of FLUX 3 Image Klein, we ran it again with the separate model its maker says to use, a prompt upsampler that rewrites a short prompt into a long caption, and the same judge, ranking blind on new pictures, put the caption's picture above the as-typed picture on 5 of 6 subjects, where the rule written first needed 4; nothing moves the game's painter.", "published": "2026-10-27", "series": [ "bench" ], "licence": "CC BY 4.0 for the words and tables, except the one caption the maker's prompt upsampler wrote, which the page carries under Apache 2.0; the six pictures are the full exhibit's own files and carry FLUX 3 Image Klein's FLUX Kommunity License v1.0 (non-commercial use), 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-with-the-prompt-upsampler/", "md_url": "https://research.strata2signal.com/flux3-klein-with-the-prompt-upsampler/index.md", "md_sha": "f0ea94c3734f4122d168fe8e9a1da173f98ccf4cdeb58705511f022bd04ab027", "html_sha": "7c6a8d2e0e8a82d24dae9313594fdef3af5903d0bf6edd20eb7d4403d21657a6", "anchors_sha": "159bffa4c93ffcee54a801e2da9766656e892f7af051d05b34dbe7ecbb2f5aa6" }, "short": { "paragraph": "A prompt upsampler is a separate model that rewrites a short request into a long caption, the kind of text the maker says the picture model was trained on. On 2026-10-24 we ran FLUX 3 Image Klein two ways on each of six real subjects from our game, on one NVIDIA GeForce RTX 3090 (24 GB) at 300 W, at 768 × 768: reading our prompt as typed, and reading the caption its maker's upsampler wrote from that same prompt. One judge, the member of our team who directs the game's art, blind to the labels, ranked the caption's picture above the as-typed picture on 5 of 6; the rule, frozen twelve and a half hours before the first picture, needed 4, and a judge ranking at random clears that bar 22 times in 64. The caption's picture also placed above FLUX.1 [dev] on 4 of 6, reported, not a rule, and the judge, asked afterwards, named the upsampled picture on 4 of 6 sheets. A win here buys a sentence on the page, never a place in the game or the print lab: nothing moves the game's painter.", "from_dek": false, "counts": { "words": 2491, "minutes": 11, "tables": 4, "kit": false }, "bullets": [ { "text": "the plan was frozen at 04:14:23Z", "figure": "04:14:23Z", "cite": "the-rule", "quote": "The rule, written down first The plan, with its predictions, was frozen at 04:14:23Z on 2026-10-24 (UTC), twelve and a half hours before the first picture, and never amended.", "span": [ 4333, 4343 ] }, { "text": "the counted run took 19 minutes 31 seconds", "figure": "19 minutes 31 seconds", "cite": "how-it-ran", "quote": "The counted run took 19 minutes 31 seconds, 16:48:21Z to 17:07:52Z on 2026-10-24 (UTC): 108 pictures and 6 captions, all valid, so the bench was COMPLETE.", "span": [ 6938, 6960 ] } ] }, "sections": [ { "id": "what-changed", "heading": "What changed, and why a caption matters", "level": 2, "span": [ 2519, 4243 ], "chunks": [ [ 2519, 4243 ] ], "chars": 1724, "digest": null, "digest_skipped": null }, { "id": "the-rule", "heading": "The rule, written down first", "level": 2, "span": [ 4243, 6614 ], "chunks": [ [ 4243, 6614 ] ], "chars": 2371, "digest": null, "digest_skipped": null }, { "id": "how-it-ran", "heading": "How it ran", "level": 2, "span": [ 6614, 7502 ], "chunks": [ [ 6614, 7502 ] ], "chars": 888, "digest": "the test ran on an NVIDIA GeForce RTX 3090 (24 GB) using ComfyUI v0.42.5 and PyTorch 2.13.0. the counted run took 19 minutes 31 seconds, from 16:48:21Z to 17:07:52Z on 2026-10-24 (UTC), producing 108 pictures and 6 captions.", "digest_skipped": null }, { "id": "what-the-judge-ranked", "heading": "What the judge ranked", "level": 2, "span": [ 7502, 9104 ], "chunks": [ [ 7502, 9104 ] ], "chars": 1602, "digest": "the judge ranked the caption's picture E above the as-typed picture C′ on 5 of the 6 subjects, resulting in a PUP-WIN. the judge also named the upsampled picture on 4 of the 6 sheets, while a random guesser names it on 4 or more in 154 of 4,096 tries.", "digest_skipped": null }, { "id": "the-pictures", "heading": "Three of the pictures", "level": 2, "span": [ 9104, 15863 ], "chunks": [ [ 9104, 15863 ] ], "chars": 6759, "digest": null, "digest_skipped": null }, { "id": "what-this-cannot-say", "heading": "What this cannot say", "level": 2, "span": [ 15863, 17350 ], "chunks": [ [ 15863, 17350 ] ], "chars": 1487, "digest": "the results are limited to one judge, one seed on the sheet, six subjects, one style, and one card. the test is not the maker's full current recipe and does not change anything for the game, as FLUX.1 [dev] stays the painter.", "digest_skipped": null }, { "id": "the-full-exhibit", "heading": "The full exhibit", "level": 2, "span": [ 17350, 18198 ], "chunks": [ [ 17350, 18198 ] ], "chars": 848, "digest": null, "digest_skipped": null }, { "id": "who-ran-this-and-thanks", "heading": "Who ran this, and thanks", "level": 2, "span": [ 18198, 18930 ], "chunks": [ [ 18198, 18930 ] ], "chars": 732, "digest": null, "digest_skipped": "credits" }, { "id": "sources", "heading": "Sources", "level": 2, "span": [ 18930, 20018 ], "chunks": [ [ 18930, 20018 ] ], "chars": 1088, "digest": "sources 1 to 4 are cited from the full exhibit, and the page adds no source of its own.", "digest_skipped": null } ], "chips": [ { "id": "c-a94db79e", "q": "how long did the counted run take to complete?", "a": "The counted run lasted 19 minutes 31 seconds, occurring between 16:48:21Z and 17:07:52Z on 2026-10-24 (UTC). This process included 108 pictures and 6 captions.", "cites": [ "how-it-ran" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-0f2c8eac", "q": "how many subjects favored the upsampled caption?", "a": "The judge ranked the caption's picture above the as-typed picture on 5 of the 6 subjects. Because the rule required at least 4 subjects to favor the caption, the result is a PUP-WIN.", "cites": [ "what-the-judge-ranked" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-78d56846", "q": "what visual differences appear in the wrenna sault portrait?", "a": "The as-typed picture depicts a young woman with dark hair and small earrings, while the caption's version shows a middle-aged woman in a dark hood with grey-streaked hair, holding a glass oil lamp and brass-handled scissors.", "cites": [ "the-pictures" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-89a6909f", "q": "what limitations exist regarding the testing scope?", "a": "The test is limited to one judge, one seed per sheet, six subjects, one style, and one card. It does not represent other upsamplers, different seeds, or the maker's full current recipe, which uses a different timestep shift and default size.", "cites": [ "what-this-cannot-say" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-2c4654fb", "q": "which documents provide the model card information?", "a": "Information regarding the model cards is found in the sources, which include the BFL FLUX 3 Image Klein model card and the BFL FLUX 3 Image Klein Prompt Upsampler model card.", "cites": [ "sources" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" } ], "related": [ { "slug": "flux3-klein-in-the-coves-style", "why": "shares ground with § What this cannot say · § What this changes for the cove" }, { "slug": "the-new-kid", "why": "shares ground with § What the judge ranked · § The four answers" }, { "slug": "outside-judges", "why": "shares ground with § What the judge ranked · § The gauntlet: our judges, examined" } ], "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThanks to Black Forest Labs for the upsampler, its system prompts and its workflow files, and to the ComfyUI team. Software this page stood on: **ComfyUI** v0.42.5 (GPL-3.0), **PyTorch** 2.13.0 (BSD-style), **CUDA** (NVIDIA, proprietary), **Hugging Face transformers** 5.20.2 (Apache 2.0), whose tokenizer counted the caption tokens, and BFL's **FLUX 3 Image Klein Prompt Upsampler** (Apache 2.0, derived from Qwen3.5-9B). Models: FLUX 3 Image Klein (FLUX Kommunity License v1.0), FLUX.1 \\[dev\\] (non-commercial), FLUX.2 \\[klein\\] 4B (Apache 2.0). A small human team set the rule, made the blind ranks and signed the numbers; AI agents ran the harness and drafted this page.", "kit": null, "seat_class": { "writer": "gemma-class", "judge": null, "writer_runtime": "vllm" }, "bench": { "bullets_written": 3, "bullets_kept": 2, "digests_written": 8, "digests_kept": 4, "chips_written": 8, "chips_kept": 5, "chips_grounded": 0, "chips_published": 0, "dropped_by": { "new_noun": 0, "figure": 0, "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": 8 }, "new_noun_tokens_checked": [ "04", "07", "096", "1", "108", "14", "154", "16", "17", "18", "19", "2.13.0", "2026-10-15", "2026-10-24", "21Z", "228", "23Z", "24", "244", "266", "3", "3090", "31", "4", "48", "5", "5.94", "52Z", "54", "6", "7.14", "768", "BFL", "C'", "ComfyUI", "E", "FLUX", "FLUX.1", "GB", "GeForce", "Image", "Klein", "Klein's", "NVIDIA", "PUP-WIN", "Prompt", "PyTorch", "RTX", "Salt-Father", "Sault", "UTC", "Upsampler", "Wrenna", "seed-404", "v0.42.5" ], "new_noun_tokens_withheld": 0, "figure_definition": "v3", "figure_boundary": "guarded", "drops": [ { "kind": "bullet", "reason": "hand-edit", "item": null, "item_chars": 75, "text_sha": "9ea67b0eb098", "figure": "5", "withheld": "one bullet removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Ends on a bare 5 that loses what the page read it against: the judge ranked the caption's picture above the as-typed picture on 5 of the 6 subjects, the rule needed 4, and the result is PUP-WIN; the line drops the six, the bar and the registered word.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Ends on a bare 5 that loses what the page read it against: the judge ranked the caption's picture above the as-typed picture on 5 of the 6 subjects, the rule needed 4, and the result is PUP-WIN; the line drops the six, the bar and the registered word. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "digest", "reason": "hand-edit", "item": null, "item_chars": 296, "text_sha": "7c0c297d77a7", "section": "what-changed", "withheld": "one section digest removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Gives the caption lengths as the upsampler's own, where the page says that on our six prompts it wrote captions of 228 to 266 words (median 244); and the fifth more per picture loses what it measures, a median 7.14 against 5.94 seconds over seeds 202 and 303.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Gives the caption lengths as the upsampler's own, where the page says that on our six prompts it wrote captions of 228 to 266 words (median 244); and the fifth more per picture loses what it measures, a median 7.14 against 5.94 seconds over seeds 202 and 303. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "digest", "reason": "hand-edit", "item": null, "item_chars": 236, "text_sha": "0b61ad8b6ea6", "section": "the-rule", "withheld": "one section digest removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Misnames the two arms and drops the bar: E is FLUX 3 Image Klein reading the upsampler's caption and C′ the same model reading the prompt as typed (the upsampler draws nothing), and PUP-WIN is E above C′ on at least 4 of the 6 subjects, read only if at least 4 subjects have a valid caption and both pictures, a tie not above.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Misnames the two arms and drops the bar: E is FLUX 3 Image Klein reading the upsampler's caption and C′ the same model reading the prompt as typed (the upsampler draws nothing), and PUP-WIN is E above C′ on at least 4 of the 6 subjects, read only if at least 4 subjects have a valid caption and both pictures, a tie not above. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "digest", "reason": "hand-edit", "item": null, "item_chars": 280, "text_sha": "69b463197f3f", "section": "the-pictures", "withheld": "one section digest removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Says the section gives the captions of the upsampled pictures, where it prints one, Wrenna Sault's (all six are on the full exhibit), and has the full exhibit show each subject's seed-404 sheet pictures, where it shows the three it may show (FLUX.1 [dev]'s are held). The section hangs six pictures, the rule's pair on three subjects.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Says the section gives the captions of the upsampled pictures, where it prints one, Wrenna Sault's (all six are on the full exhibit), and has the full exhibit show each subject's seed-404 sheet pictures, where it shows the three it may show (FLUX.1 [dev]'s are held). The section hangs six pictures, the rule's pair on three subjects. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "digest", "reason": "hand-edit", "item": null, "item_chars": 235, "text_sha": "0564c2e44e0a", "section": "the-full-exhibit", "withheld": "one section digest removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Reads as if the run of 2026-10-15 made 54 pictures: the page says the full exhibit shows the 54 that FLUX 3 and klein 4B made, and FLUX.1 [dev]'s pictures from that run are held.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Reads as if the run of 2026-10-15 made 54 pictures: the page says the full exhibit shows the 54 that FLUX 3 and klein 4B made, and FLUX.1 [dev]'s pictures from that run are held. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "chip", "reason": "hand-edit", "item": null, "item_chars": 304, "text_sha": "67d7d783c04d", "chip_id": "c-f3d3d361", "withheld": "one chip removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Gives the six captions' lengths as the upsampler's own (the page: on our six prompts, 228 to 266 words, median 244) and the seconds without their seeds: the median 7.14 against 5.94 seconds is over seeds 202 and 303.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Gives the six captions' lengths as the upsampler's own (the page: on our six prompts, 228 to 266 words, median 244) and the seconds without their seeds: the median 7.14 against 5.94 seconds is over seeds 202 and 303. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "chip", "reason": "hand-edit", "item": null, "item_chars": 259, "text_sha": "3ed6d5325562", "chip_id": "c-49f25bfb", "withheld": "one chip removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Weakens the condition the rule is read under: the page says read only if at least 4 subjects have a valid caption and both pictures, and a tie is not above; the answer asks for one picture and drops the tie.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Weakens the condition the rule is read under: the page says read only if at least 4 subjects have a valid caption and both pictures, and a tie is not above; the answer asks for one picture and drops the tie. Not a gate drop, and recorded here so the ledger adds up." }, { "kind": "chip", "reason": "hand-edit", "item": null, "item_chars": 174, "text_sha": "4d47658722d0", "chip_id": "c-03dda7c0", "withheld": "one chip removed from this pack by hand", "removed_utc": "2026-10-27", "why": "Reads as if the run of 2026-10-15 made 54 pictures: the page says the full exhibit shows the 54 that FLUX 3 and klein 4B made, and FLUX.1 [dev]'s pictures from that run are held.", "detail": "removed from this pack by hand on 2026-10-27 (UTC), after the writer ran and after the gates had counted: Reads as if the run of 2026-10-15 made 54 pictures: the page says the full exhibit shows the 54 that FLUX 3 and klein 4B made, and FLUX.1 [dev]'s pictures from that run are held. Not a gate drop, and recorded here so the ledger adds up." } ] }, "generated_utc": "2026-10-27T00:10:12Z" }