The bench
FLUX 3 Image Klein, ranked blind in a game's own painted style
exhibit eighty-eight The bench
Published 2026-10-20 (UTC)
updated 2026-10-26 (UTC)
A small (human) team and a fleet of AI agents.
Black Forest Labs released the image model FLUX 3 Image Klein on 2026-10-15. We ran it that day on six real subjects from our game; one person, ranking blind, put it above FLUX.1 [dev], the game's painter, on 3 of the 6 and below it on 3, one short of the bar we set first.
ask about this page → assistant.strata2signal.com · in beta, still being tested
hardware on this page: RTX 3090 (24 GB) → research.strata2signal.com/hardware/ · the roster is in beta, still being tested
the short version
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.
7,495 words, about 34 minutes to read.
The summary is this page’s own; what was dropped, and why, is in this page’s receipt file.
If you're new here: strata→signal is a small workshop that runs its own machines and writes up what it measures. Web pages cited were read on 2026-10-15 (UTC).
What the model is
FLUX 3 Image Klein turns a sentence into a picture, with 5.6 billion parameters plus a separate text encoder [2]. Black Forest Labs (BFL) released it as open weights, downloadable after sharing contact details on Hugging Face [1][2].
- Steps. It cleans a picture out of noise a step at a time. We ran the distilled version, trained to be quick: "16 steps and no guidance" [3] (guidance pushes the picture harder toward the prompt; cfg 1 means none).
- Licence. The FLUX Kommunity License v1.0: free for non-commercial use; people and companies under US$5 million a year in revenue, affiliates included, may also use its pictures commercially, subject to its terms [1][2].
- The maker's claim. "Compared with FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image" [1]. Elo is a rating scale first built for chess: the gap between two ratings predicts how often the higher one wins, and 91 points is about 63 wins in 100 head-to-head votes. The post gives no vote counts, voters or prompts, and our klein column is FLUX.2 [klein] 4B, not the 9B.
What we tested
RealKeep, our small role-playing game, paints sketches of its fishing-village world, Sorrowmoor Cove ("the cove"), in one house style [7]. The question: did one judge, looking blind at one picture per subject, rank FLUX 3 Image Klein above FLUX.1 [dev] on at least 4 of the game's 6 real subjects in the cove's style, and at what cost in seconds and memory?
| Column | Model and build | Steps | Licence |
|---|---|---|---|
| A | FLUX.1 [dev], 8-bit floating-point (fp8) build | 32, guidance 3.5: the game's recipe | non-commercial [5] |
| B | FLUX.2 [klein] 4B, BF16 | 8, cfg 1 (our earlier bake-off's setting; the maker's example: 4 [6]) | Apache 2.0 [6] |
| C | FLUX 3 Image Klein, distilled, BF16 | 16, no guidance | FLUX Kommunity v1.0 |
| D | the same, in the int8 build for "GPUs without FP8 support (RTX 30 series and older)" [3] | 16, no guidance | the same |
The game runs FLUX.1 [dev] in 16-bit; the bench computer holds only the fp8 build, so column A is not the game's exact build. klein 4B paints for our print lab. All four ran in ComfyUI, a free picture-making program, at v0.41.0, the first version that runs FLUX 3 Image [4]. The prompt upsampler, a model that rewrites short prompts [3], was off, so all four read the same words.
The style, and the six subjects
The style, verbatim from the game's campaign file:
stylized oil-painted realism, loose quick brushwork, in the style of Dishonored game art, unsigned, no text, no letters, no signature, no watermark
Dishonored is a 2012 Arkane Studios game [8], named only for its style; no picture from it went to any model. Each prompt is the style clause, a comma, then one subject, unedited, with no negative prompt.
The six subjects, and the judge's check list, fixed in advance:
| Subject | Clause, verbatim | What must show |
|---|---|---|
| Foe · the Reef-Runner | a hard-faced smuggler in dripping oilskins wading a flooded sea-cave, a gaff hook in one hand and a hooded lantern in the other | a gaff hook · a hooded lantern |
| Foe · the Salt-Father | a towering humanoid monster built of jagged translucent white salt crystal, looming full-figure, smaller dry husk-people half-merged into its glittering salt mass, no face | built of salt crystal · no face |
| Scene · the sea-fog breakwater | REGION VIGNETTE: A thick curtain of sea-fog swallowing the stone breakwater and the distant silhouette of forty huddled rooftops. | the stone breakwater · huddled rooftops in the distance |
| Scene · the stone chapel | REGION VIGNETTE: The weathered stone chapel with its cracked bell tower glowing under a dim lantern light. | a cracked bell tower |
| Portrait · Harbormaster Haldane | close-up head-and-shoulders portrait of a precise middle-aged harbormaster in a peaked cap, skeptical fair eyes, his pen poised mid-entry, the ledger counter a blur below, amber on grey | a peaked cap · a pen |
| Portrait · Wrenna Sault | close-up head-and-shoulders portrait of a quiet devout lampwright, wick-trimming scissors in hand, warm amber glow on her face, night fog-blue behind | wick-trimming scissors |
The rules, written down first
Registered at 19:09:33Z on 2026-10-15, before the first picture (a smoke test's uncounted warm-up, finished at 21:32:04Z); amended at 19:46:43Z (a faster out-of-memory stop) and, before the key was opened, at 01:12:01Z on 2026-10-16 (ties allowed).
- Valid? Each picture needs the registered model, steps, size and prompt, nothing else on the card, and no model "loaded partially" (part left in main memory).
- The blind rank. Each subject's four seed-101 pictures (a seed fixes the starting noise) go on one sheet, shuffled and unlabelled, the key sealed. The judge, the member of our team who directs the game's art, ranks them 1 (best) to 4: "which best reads as the cove's style … while showing the subject named above?", and ticks the "what must show" items.
- A WIN is column C ranked above column A on at least 4 of the 6 subjects, read only if at least 4 subjects have a valid picture from both (else INCOMPLETE). A WIN buys a trial, never a place; a NO-WIN closes this line. klein 4B, int8, the ticks and the seconds decide nothing.
Four of six is a modest bar: a judge with no preference would clear it in 22 of every 64 tries.
How it ran
The NVIDIA GeForce RTX 3090 had nothing else on it; its power limit read 300 W before every column, never changed. The run took 13 minutes 44 seconds, 21:33:00Z to 21:46:44Z on 2026-10-15 (UTC). A smoke test ran just before it, 21:31:22Z to 21:32:11Z: column C's warm-up, then one Reef-Runner picture at seed 101, to prove the path; neither picture is one of the run's 72. Each column ran on a fresh ComfyUI, in the order C, A, D, B: an uncounted warm-up, then the six subjects at seeds 101, 202 and 303, at 768 × 768. All 72 pictures were valid, so the bench was COMPLETE. The card peaked at 74 °C (the stop: 83 °C).
Memory. Both FLUX 3 builds fit whole on the card at 768 × 768, under ComfyUI's older loader (dynamic VRAM off), which logs whether each part loaded whole: no picture in any column loaded partially. Memory peaked at 21,148 MiB of 24,576 for BF16, 12,306 for int8.
Seconds per picture
ComfyUI's clock covers steps, decoding, saving and, on each subject's first seed, reading the prompt, so the fair line is seeds 202 and 303.
| Column | Steps | Median seconds, seeds 202 + 303 (12 pictures) | Range of the 12 | Warm-up (loading included) |
|---|---|---|---|---|
| A · FLUX.1 [dev], fp8 | 32 | 22.77 | 22.63–22.79 | 68.26 |
| B · FLUX.2 [klein] 4B, BF16 | 8 | 2.84 | 2.82–2.90 | 19.55 |
| C · FLUX 3 Image Klein, BF16 | 16 | 5.87 | 5.72–6.16 | 14.23 |
| D · FLUX 3 Image Klein, int8 | 16 | 2.73 | 2.71–2.91 | 13.20 |
Same run, card and ComfyUI version, so these compare directly: int8 took less than half BF16's time and, at 16 steps, about as long as klein 4B at 8; BF16 took about a quarter of FLUX.1 [dev]'s time (16 steps against 32).
What the judge ranked
The judge ranked FLUX 3 Image Klein BF16 (C) above FLUX.1 [dev] (A) on 3 of the 6 subjects; the rule needed 4, so on a COMPLETE bake-off the result is NO-WIN.
| Subject | 1st | 2nd | 3rd | 4th | C above A? |
|---|---|---|---|---|---|
| Foe · the Reef-Runner | FLUX.1 [dev] | FLUX 3 BF16 | FLUX 3 int8 | klein 4B | no |
| Foe · the Salt-Father | FLUX 3 int8 | FLUX 3 BF16 | klein 4B | FLUX.1 [dev] | yes |
| Scene · the sea-fog breakwater | FLUX 3 int8 | FLUX 3 BF16 | FLUX.1 [dev] | klein 4B | yes |
| Scene · the stone chapel | FLUX 3 BF16 | FLUX 3 int8 | klein 4B | FLUX.1 [dev] | yes |
| Portrait · Harbormaster Haldane | FLUX.1 [dev] | klein 4B | FLUX 3 BF16 | FLUX 3 int8 | no |
| Portrait · Wrenna Sault | klein 4B | FLUX.1 [dev] | FLUX 3 BF16 | FLUX 3 int8 | no |
| Of 6 | 3 |
Reported, not registered (none of it is a claim):
- FLUX 3's two builds took the top two places on both scenes and the Salt-Father, the bottom two on both portraits.
- FLUX 3 BF16 placed above klein 4B on 4 of 6; no bar was set for this pair.
- The judge ticked every "what must show" item on every ranked (seed-101) picture except "no face" on FLUX.1 [dev]'s and klein 4B's Salt-Father.
int8 against BF16 (reported, no claim). int8 placed above BF16 on 2 of 6 sheets (the Salt-Father, the breakwater) and below on 4. Two sheets first came back with shared places; the judge, still blind, wrote: "the reason i ranked some the same #s is because the images looked identical to me." Ties were then allowed (see the rules); the six final orders used none. Asked after the reveal, on the labelled pictures, whether the two builds differ, the judge answered "pretty much indistinguishable": the same. Measured on the pictures shown below, an int8 picture's pixels sit on average 1.9 to 7.9 brightness steps (out of 255) from the BF16 picture's at the same subject and seed, most on Wrenna Sault at seed 303, whose face changes; klein 4B's sit 25 to 50 steps from BF16's.
The pictures
Every picture klein 4B (column B) and FLUX 3 Image Klein (column C, BF16, and column D, int8) made is below, 54 in all: one table per subject, one row per seed. Each cell is a small copy; select it to see the 768 × 768 picture as the model drew it, saved as a JPEG. Under each is its time in seconds by ComfyUI's clock, which on seed 101 includes reading the prompt; each column's head gives its median over seeds 202 and 303, from the seconds table above. Every prompt is the style clause, a comma and the subject's clause, as printed above, with no negative prompt. Seed 101 is the picture the judge ranked blind.
FLUX.1 [dev]'s pictures (column A) are not shown: a ruling of ours from 2026-09-04, recorded in the licence ledger (linked below), holds them, so column A appears on this page in words only, in the ranks above. FLUX 3's pictures appear under the FLUX Kommunity License v1.0's non-commercial use (this site charges nothing and carries no ads); klein 4B's under Apache 2.0.
One fault the rules did not ask about shows in the corners. Every prompt says "unsigned, no text, no letters, no signature, no watermark", yet 10 of the 54 pictures carry a scrawl or a stamp like a signature in a bottom corner: 8 of klein 4B's 18 (the Salt-Father and the breakwater at all three seeds, the Reef-Runner and the chapel at seed 303) and FLUX 3's Reef-Runner at seed 101, in both builds. We counted them by eye on the pictures shown here.
Foe · the Reef-Runner
Foe · the Salt-Father
Scene · the sea-fog breakwater
Scene · the stone chapel
Portrait · Harbormaster Haldane
Portrait · Wrenna Sault
With the labels showing
This section is the judge's own look at the labelled pictures, on 2026-10-18 (UTC), two days after the blind ranks were scored. The judge could see which model made each picture, so this is an impression, not a ranking, and it changes none of the results above.
- Overall. "I pretty much prefer FLUX.2 klein 4B · 8 steps, across the board, with a few exceptions", and "but it's by far the most consistent".
- The exception named. "salt father is the exception where flux3.klein draws a closer representation of what I feel like this foe would look like". The blind ranks agree on that subject: FLUX 3's two builds placed first and second.
- Where the two reads part. Blind, the judge put FLUX 3's two builds first and second on both scenes, and above klein 4B on 4 of the 6 subjects; klein 4B placed above both only on the two portraits.
- Hands. "look at the hands". On Wrenna Sault at seed 303, both FLUX 3 builds draw the hand on the scissors as a fused lump, its fingers merged and a handle ring passing through it. At seed 101 both draw the scissors across her collar with no hand to hold them, only shapeless patches inside the handle rings at the bottom edge. Harbormaster Haldane's hands hold together in both builds at all three seeds, at this size. klein 4B draws well-formed hands in all six portrait pictures.
What this changes for the cove
Nothing. By the rule fixed before the run, FLUX.1 [dev] stays the game's painter and this line of testing closes. FLUX 3's split (top two on both scenes, bottom two on both portraits) and int8's 2.73 seconds a picture are noted for a new, separately registered test, only if the team asks for one.
What this result does and does not mean
- It is one judge's blind first look: six subjects, one style, one picture each, one card.
- Not the game's own 16-bit FLUX.1 [dev]; not BFL's 91 Elo; not the base version, the prompt upsampler, editing, other styles, sizes, seeds or cards. Two portrait prompts keep the game's colour words, in every column.
- FLUX.1 [dev] and FLUX 3 Image Klein ran as an evaluation, which we read as non-commercial use under their licences; nothing they made feeds the game or any other model.
How to check our work
No data kit ships (the run's files name our machines); the prompts, settings, rules, timings and pictures above are what a repeat needs.
- BFL, "FLUX 3 Image Klein", 2026-10-15. https://bfl.ai/blog/flux-3-image-klein
- BFL, FLUX 3 Image Klein model card. https://huggingface.co/black-forest-labs/flux-3-image-klein
- BFL, FLUX 3 Image Klein for ComfyUI model card. https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy
- ComfyUI pull request #17064 and tag v0.41.0, 2026-10-15. https://github.com/Comfy-Org/ComfyUI/pull/17064 · https://github.com/Comfy-Org/ComfyUI/tree/v0.41.0
- BFL, FLUX.1 [dev] model card. https://huggingface.co/black-forest-labs/FLUX.1-dev
- BFL, FLUX.2 [klein] 4B model card. https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
- RealKeep. https://realkeep.strata2signal.com/
- Wikipedia, "Dishonored". https://en.wikipedia.org/wiki/Dishonored
- Who makes FLUX: Black Forest Labs and its licences.
- The licence ledger: every model we run, with its licence as we read it.
Who ran this, and thanks
Thanks 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.
Corrections and later measurements will be added below, each dated (UTC), saying in plain words what it counts.
Measured again, 2026-10-24: FLUX 3 Image Klein with its maker's prompt upsampler
What this section is, for a reader who scrolled straight to it. It counts one judge's blind ranks of pictures FLUX 3 Image Klein made from captions written by its maker's prompt upsampler, against pictures it made from our prompts as typed, on the six subjects from our game, RealKeep, listed above. The pictures were made on 2026-10-24 (UTC), 16:48:21Z to 17:07:52Z; the judge's rank lines came in between 23:14:29Z and 23:15:24Z that day, and the sealed key was opened at 02:34:26Z on 2026-10-25 (UTC), after the judge's guesses. Against the first reading: on 2026-10-15, upsampler off, the judge put FLUX 3 Image Klein above FLUX.1 [dev] on 3 of 6 subjects, one short of the bar; on 2026-10-24 the same judge put its upsampled pictures above its as-typed ones on 5 of 6, where the bar was 4, and above FLUX.1 [dev] on 4 of 6 (reported, not a rule). No figure above changes, FLUX.1 [dev] stays the game's painter, and this section went up on 2026-10-26 (UTC).
Why it ran again
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. Black Forest Labs (BFL) released one with FLUX 3 Image Klein: a full fine-tune of Qwen3.5-9B, a nine-billion-parameter model that reads text and pictures, under Apache 2.0, 18.8 GB in 16-bit (BF16) [9]. On release day BFL's reference code [11] and its own ComfyUI workflow ran it by default, and its model card [3] said a short prompt gives generic output; we turned it off by our registered rule, so that all four models read the same words. From about 22:58Z on 2026-10-15, 71 minutes after our last picture, BFL's pages say "Do not use the model without a prompt upsampler" [2][3]; on 2026-10-17 its new workflows moved the timestep shift (how many of the 16 steps are spent at high noise) from α 6.93 to α 16, and the default size to 4 megapixels [10]. The weights we ran are the same bytes BFL now offers under new names, and no ComfyUI bug fix since v0.41.0 is on the path we ran (the int8 path was reworked for speed). So the ranks above stand for prompts as typed and may understate FLUX 3; this section is the upsampler reading the page above said it did not cover.
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.
| Arm | Model and build | What it read | Seeds | Role |
|---|---|---|---|---|
| C′ | FLUX 3 Image Klein, distilled, BF16; the same files and ComfyUI graph as column C above | our prompt as typed: the style clause, a comma, the subject's clause | 101, 202, 303, 404 | the control; the rule's baseline |
| E | the same model, files and graph; only the text differs | the subject's caption, written by BFL's BF16 upsampler from that same typed prompt | 101, 202, 303, 404 | the ask; the rule |
| A | FLUX.1 [dev], fp8 build, 32 steps, guidance 3.5: the game's recipe | typed | 404 | on the sheet; reported |
| B | FLUX.2 [klein] 4B, BF16, 8 steps | typed | 404 | on the sheet; reported |
| D′ | FLUX 3 Image Klein, int8 build, the same as column D above | typed | 101, 202, 303, 404 | reported only; not on the sheet |
| F | E with the maker's newer timestep shift, α 16, at our size | the caption | 101, 202, 303, 404 | reported only; never on the sheet |
- The caption. The upsampler ran alone on the card first, with BFL's own workflow settings [10]: its text-to-image system prompt (standing instructions read before every request; 3,404 characters [9]), greedy decoding (always the top-scoring next token, a word or part of one, so a request gets the same caption at every seed) and thinking off. It read each typed prompt byte for byte; the six captions were saved and checksummed before the first E picture, and none was re-rolled or edited.
- The sheet. One per subject: the seed-404 pictures of A, B, C′ and E, shuffled, unlabelled, resampled to 768 pixels and stripped of metadata, the key sealed and its checksum printed on every sheet. Seed 404 is fresh: the judge, the member of our team who directs the game's art, had already seen every picture at seeds 101, 202 and 303 labelled, FLUX.1 [dev]'s included (the labelled look above is dated 2026-10-18). The sheet shows the typed prompt, never a caption, with the question and "what must show" ticks above.
- PUP-WIN (PUP: prompt upsampler) is E ranked 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 is not above. A judge ranking at random clears it 22 times in 64. It buys a sentence on this page, never a place in the game or the print lab, and the line closed on 2026-10-15 stays closed. A, B, D′, F, the ticks, the seconds and the guess decide nothing.
- The guess, after all six ranks, with the key still sealed: on each sheet the judge named the picture taken to be the upsampled one. It changes no rank. A random guesser names it on 4 or more of 6 sheets in 154 of 4,096 tries.
- A stop. Had any as-typed BF16 picture at seeds 101, 202 or 303 sat more than 8.0 brightness steps (out of 255) from its lossless original above, the run would have stopped before any ranking.
How it ran
It ran on one NVIDIA GeForce RTX 3090 (24 GB), the same board as above, alone on the card, its power limit read at 300 W before every arm; its highest readings were 69 °C and 291 W. ComfyUI was v0.42.5 (v0.41.0 above), PyTorch 2.13.0, dynamic VRAM off, no custom nodes, offline, every model loading whole. After an uncounted dry run (16:44:58Z to 16:48:20Z), the counted run took 19 minutes 31 seconds, 16:48:21Z to 17:07:52Z on 2026-10-24 (UTC), each arm on a fresh ComfyUI in the order C′, captions, E, A, B, D′, F: 108 pictures and 6 captions, all valid, so the bench was COMPLETE.
The published pictures reproduce. On the newer ComfyUI the 36 as-typed FLUX 3 pictures at seeds 101, 202 and 303 (BF16 and int8) came out pixel for pixel the same as the lossless originals of the 36 above (0.000 brightness steps apart on every one). Memory peaked at 21,148 MiB of 24,576 for the BF16 arms (as above) and 12,284 MiB for int8 (12,306 above); the upsampler's 17,948 MiB of weights loaded whole, alone.
The captions ran 228 to 266 words (median 244), or 293 to 357 tokens. The first took 24.49 seconds by ComfyUI's clock, the upsampler's 18.8 GB load included; the others 7.33 to 8.83 seconds, about 40 tokens a second. All six keep "Dishonored", "oil-painted" and "brushwork", and none asks for a photograph or quotes text to draw. Five restate "no signature, no watermark" as what is present ("free of typography or graphic marks"); the Salt-Father's keeps the words ("unmarked by typography, emblems, signatures, or watermarks").
Seconds per picture, by ComfyUI's clock: for C′, E, D′ and F the median of seeds 202 and 303 (12 pictures), as above; A and B ran seed 404 only, so their six include reading the prompt.
| Arm | Steps | Median seconds | Range | Against 2026-10-15 (no claim: a newer ComfyUI) |
|---|---|---|---|---|
| C′ · FLUX 3 Image Klein BF16, typed | 16 | 5.94 | 5.82–6.17 | 5.87 |
| E · the same, reading the caption | 16 | 7.14 | 6.98–7.36 | — |
| F · E at α 16 | 16 | 7.12 | 6.98–7.37 | — |
| D′ · FLUX 3 Image Klein int8, typed | 16 | 2.71 | 2.66–2.87 | 2.73 |
| A · FLUX.1 [dev] fp8 (six seed-404 pictures, prompt reading included) | 32 | 22.85 | 22.45–22.95 | 22.77 (seeds 202 + 303) |
| B · FLUX.2 [klein] 4B (six seed-404 pictures, prompt reading included) | 8 | 3.10 | 2.81–3.18 | 2.84 (seeds 202 + 303) |
The caption costs about a fifth more per picture (7.14 against 5.94 seconds); the newer shift costs nothing (7.12 against 7.14 seconds).
What the judge ranked
The judge ranked the caption's picture (E) above the as-typed picture (C′) on 5 of the 6 subjects; the rule needed 4, so on a COMPLETE sheet the result is PUP-WIN. A judge ranking at random clears that bar 22 times in 64.
| Subject | 1st | 2nd | 3rd | 4th | E above C′? | E above FLUX.1 [dev]? |
|---|---|---|---|---|---|---|
| Foe · the Reef-Runner | FLUX.1 [dev] | FLUX 3, caption | FLUX 3, typed | klein 4B | yes | no |
| Foe · the Salt-Father | FLUX 3, caption | FLUX 3, typed | klein 4B | FLUX.1 [dev] | yes | yes |
| Scene · the sea-fog breakwater | FLUX 3, typed | FLUX 3, caption | FLUX.1 [dev] | klein 4B | no | yes |
| Scene · the stone chapel | FLUX.1 [dev] | klein 4B | FLUX 3, caption | FLUX 3, typed | yes | no |
| Portrait · Harbormaster Haldane | FLUX 3, caption | FLUX.1 [dev] | FLUX 3, typed | klein 4B | yes | yes |
| Portrait · Wrenna Sault | FLUX 3, caption | klein 4B | FLUX.1 [dev] | FLUX 3, typed | yes | yes |
| Of 6 | 5 | 4 |
The guess. The judge named the upsampled picture on 4 of the 6 sheets (the Salt-Father, the chapel, Harbormaster Haldane, Wrenna Sault); a random guesser names it on 4 or more in 154 of 4,096 tries. On the two misses, the Reef-Runner and the breakwater, the guess fell on the as-typed FLUX 3 picture, so every guess landed on FLUX 3.
Reported beside the rule (none of it decides anything or is a claim):
- E placed above FLUX.1 [dev] on 4 of 6 and above klein 4B on 5 of 6 (all but the chapel); no bar was set for either. The as-typed picture placed above FLUX.1 [dev] on 2 of 6 (the Salt-Father, the breakwater), against 3 of 6 on 2026-10-15 at seed 101: a different seed, so no claim.
- No ties were written. Every "what must show" item was ticked on all six E and all six C′ pictures (10 of 10 each). FLUX.1 [dev] missed "no face" on the Salt-Father and the scissors on Wrenna Sault, and klein 4B missed "no face" on the Salt-Father; both missed that one on 2026-10-15 too.
- F, the caption at the newer shift α 16, was measured, never ranked: each F picture sits on average 6.5 to 14.7 brightness steps from E's at the same subject and seed (median 10.1, 24 pairs). On the same measure, made on the lossless originals, int8 against BF16 as typed sits 1.5 to 7.6 brightness steps (median 2.3, 24 pairs), the range those originals gave on 2026-10-15; the 1.9 to 7.9 above was measured on the JPEGs shown. So the shift's median move is larger than any move int8 made.
The pictures
One table per subject below holds the three seed-404 sheet pictures this page may show: C′ (as typed), E (the caption) and klein 4B (B); FLUX.1 [dev]'s are held by the ruling of 2026-09-04 and appear in words only. Each cell is a small copy; select it to see the 768 × 768 picture as the model drew it, saved as a JPEG. Under each is its time in seconds by ComfyUI's clock (klein 4B's includes reading the prompt). E's caption is printed in full under each table; C′ and B read the typed prompt in the subjects table above. FLUX 3 Image Klein's pictures appear under the FLUX Kommunity License v1.0's non-commercial use (this site charges nothing and carries no ads); klein 4B's and the captions under Apache 2.0.
Counted by eye on these 18: none of E's six carries a scrawl like a signature, and one of C′'s does (the Salt-Father's, bottom centre). Of klein 4B's six, one carries a signature-like scrawl (the Salt-Father's, bottom right) and two carry letter-like marks elsewhere (Harbormaster Haldane's ledger page, a patch on the Reef-Runner's chest).
Foe · the Reef-Runner, seed 404
The caption E read (263 words, the upsampler's output from the typed prompt): Stylized oil-painted realism with loose, quick brushwork and the dark fantasy game-art finish of Dishonored, showing a flooded sea-cave from a low three-quarter viewpoint. In the center, a hard-faced male smuggler in his forties wades through waist-deep black water, his broad shoulders pitched forward against the current. He has weathered olive skin, a square jaw, heavy brow, short dark stubble, and narrowed eyes beneath a deep hood. Thick rivulets run down his cheeks and drip from the edges of his battered, oil-soaked leather clothing. His long green-black oilskins are saturated with seawater, their creases streaked by dripping oil that falls in irregular reflective trails. A frayed rope belt circles his waist, and one torn cuff exposes a dark wool sleeve. In his raised right hand he grips a long iron gaff hook, its crooked metal tip cutting diagonally toward the upper left above the water. In his left hand he carries a dented brass lantern wrapped in a soaked brown hood, casting a concentrated amber light across his face, hands, and the nearest spray. The cave walls curve around him as jagged rock formations painted in layered blue-grey, umber, and muted teal strokes. Pale shafts of lantern light reveal suspended bubbles, drifting silt, broken shells, and strands of kelp sweeping past his legs. White foam breaks around his body in broad, agitated brushstrokes. Behind him, the cavern recedes into dark arches and indistinct rock silhouettes, while scattered points of reflected amber light tremble across the flooded surface. The entire composition remains purely pictorial, with uninterrupted painted surfaces free of typography or graphic marks.
Foe · the Salt-Father, seed 404
The caption E read (266 words): Stylized oil-painted realism with loose, quick brushwork and the dark gothic fantasy visual language of Dishonored game art. A towering humanoid monster built entirely from jagged translucent white salt crystal looms in a full-body view at the center, rising from the lower foreground to near the top of the composition. Its immense body has broad shoulders, long angular arms, an oversized chest, and tapering crystalline legs planted among broken slabs of pale ground. Sharp facets split into fractured planes across its torso and limbs, catching cold blue-white highlights while deep grey shadows reveal the thickness of the crystal structure. The head is a smooth, featureless mass of faceted salt with no eyes, nose, mouth, ears, or face; only irregular ridges and splintered edges define the skull-like crown. Smaller dry husk-people are half-merged throughout the glittering salt mass: several skeletal human forms cling to its left shoulder and ribs, their cracked brown shells dissolving into white crystals, while others emerge from the right side as shriveled torsos and thin limbs trapped beneath transparent shards. Their empty sockets and exposed bones remain visible through the crystal. In the foreground, scattered salt fragments, dead reeds, and curled husks surround the monster’s feet. Behind it, a barren coastal wasteland recedes into smoky charcoal clouds and muted ochre haze, with distant ruined arches reduced to rough silhouettes. Cold light breaks through the upper background, illuminating the salt edges and leaving the outer corners in heavy shadow. Broad, hurried strokes, scraped paint texture, restrained blue-grey, umber, bone white, and tarnished gold palette; every surface remains unmarked by typography, emblems, signatures, or watermarks.
Scene · the sea-fog breakwater, seed 404
The caption E read (228 words): Stylized oil-painted realism with loose, quick brushwork in the style of Dishonored game art, composed as a wide coastal vignette. A thick curtain of blue-grey sea fog fills most of the scene, swallowing the horizon and curling around a massive stone breakwater that disappears into the mist near the center-left. The breakwater begins as a dark, jagged mass in the lower foreground, its wet blocks suggested by broad strokes of charcoal, slate blue, and dirty white before dissolving completely into the fog. On the right, forty huddled rooftops form one distant silhouette along the horizon, compressed into a low cluster of steep roofs, chimneys, and clustered parapets; only their collective outline can be read through the haze, with individual buildings lost inside the dense vapor. Pale grey-green waves surge across the bottom, breaking into rough streaks of foam that vanish beneath the fog curtain. Cold overcast light filters through the upper half as layered veils of pearl grey, muted teal, and bruised navy paint. Thick dragged bristle marks, dry-brushed edges, scraped pigment, and irregular patches of exposed underpainting give the sea and mist an urgent hand-painted texture. The outer corners fall into deep shadow while the central fog remains opaque and dominant, erasing all architectural detail beyond the breakwater and rooftop silhouette. The entire surface contains only painted seascape forms, with clean unmarked edges and no graphic overlays.
Scene · the stone chapel, seed 404
The caption E read (235 words): Stylized oil-painted realism with loose, quick brushwork in the style of Dishonored game art, composed as a region vignette. A weathered stone chapel stands alone against a dark, indistinct night landscape, viewed from a low three-quarter angle. The chapel fills the center and lower half of the composition, its squat nave built from uneven grey-brown masonry with chipped edges, patches of moss, and deep mortar seams. Above it rises a narrow bell tower split by branching cracks; one side of the tower has broken away to expose jagged stones, while an open arched belfry glows under dim amber lantern light. Inside the belfry, a single old iron lantern hangs from a crooked bracket, casting warm orange light across the cracked arches and leaving the upper tower in brown-black shadow. A small bronze bell rests within the opening, partly obscured by darkness. On the left wall, a recessed wooden door sits beneath a shallow stone arch, reached by two worn steps scattered with wet leaves and fragments of rubble. Thin smoke curls around the lantern and catches the amber illumination along the chapel’s roofline. The outer corners dissolve into rough, irregular brushstrokes that fade into near-black umber, forming a painted vignette border without a hard outline. Cool blue-grey shadows surround the structure, while the lantern creates restrained ochre highlights on nearby stones. Every surface remains unmarked, with only stone, wood, foliage, smoke, and paint texture present.
Portrait · Harbormaster Haldane, seed 404
The caption E read (242 words): Stylized oil-painted realism with loose, quick brushwork and the weathered maritime character design of Dishonored game art, a close-up head-and-shoulders portrait dominated by an amber-and-grey palette. Centered in the frame is a precise middle-aged white male harbormaster with fair skin, a broad nose, grey eyebrows, and skeptical pale eyes fixed on the viewer beneath lowered brows. His mouth rests in a firm, questioning line, one corner turned down as if he has just rejected what he heard. He wears a dark charcoal peaked harbor cap with a worn amber band, a short brim, and a small plain metal badge without markings. Wind-tangled grey hair shows at his temples above a trimmed moustache and short beard. A heavy navy wool coat with a high collar covers his shoulders. In front of his chest, his right hand holds a dark fountain pen poised mid-entry, its nib hovering above the unseen ledger surface while his gaze remains directed toward the viewer. Below him, the ledger counter dissolves into a broad blurred band of ochre paper, dark wood, and indistinct page edges. Behind his head, rough harbour walls, rigging, and cold overcast water are reduced to broken grey strokes with muted amber light catching one side of his face and cap. Thick impasto marks define the eyes, nose, and moustache, while the outer brushwork remains fast and economical. The painted surface contains only the portrait and unmarked harbor forms, with clean uninterrupted paint around the edges.
Portrait · Wrenna Sault, seed 404
The caption E read (246 words): Stylized oil-painted realism with loose, quick brushwork and the dark fantasy game-art character of Dishonored, a close-up head-and-shoulders portrait of a quiet devout lampwright against a fog-blue night background. Centered in the foreground is a white woman in her forties with fair skin, a narrow face, grey eyes, and dark brown hair threaded with silver, gathered into a low bun beneath a worn charcoal wool hood. Her closed mouth and lowered gaze rest on a small brass wick-trimming scissors held near her chest. The scissors have slender steel blades and one blade closes around the blackened tip of a cotton wick rising from an unlit glass oil lamp just below her chin. She wears a plain iron devotional cross pendant over a soot-marked umber work smock, its collar fastened with a dull brass clasp. In her raised right hand, she holds the open wick-trimming scissors beside the flame, while her left hand steadies the lamp’s tarnished brass socket. A warm amber flame illuminates one side of her face, catching her brow, cheek, nose, knuckles, and the edges of the scissors; the opposite side falls into deep brown shadow. Broad ochre and burnt-orange strokes form the glow across her features, with visible bristle marks and broken edges. Behind her, blue-grey night fog fills the upper background in broad indigo and slate strokes, with vague silhouettes of crooked rooftops and hanging lantern chains dissolving into mist. The painted surface remains entirely pictorial, with uninterrupted color and texture throughout.
The predictions, registered before the run
Twenty-three predictions were frozen with the rule at 04:14:23Z on 2026-10-24 (UTC): 22 held, 1 was refuted, and none is a gate. Each reading is from the run's own files, at the file and field the frozen plan names (unpublished: it names our machines).
| Prediction | Bar | Read | Result |
|---|---|---|---|
| The as-typed BF16 pictures replay the published ones | within 1.5 brightness steps on all 18 | 0.000 brightness steps on all 18, pixel-identical | held |
| The int8 path moved more than the BF16 path on the newer ComfyUI | int8's median drift above BF16's | both 0.000, pixel-identical | refuted (neither moved at all) |
| int8 still tracks BF16 | median int8-against-BF16 gap 1.5 to 3.0 brightness steps | 2.3 brightness steps, over 24 pairs | held |
| The BF16 upsampler loads whole, alone, on the card | its load line says so | "17947.79 MB loaded, full load: True" | held |
| The caption costs little memory | E's peak at most C′'s + 512 MiB, every load whole | 21,148 MiB for both | held |
| All six captions pass their checks | 6 of 6 | 6 of 6 | held |
| Caption length | median 150 to 300 words | 244 | held |
| The style survives | every caption keeps "Dishonored" and "oil-paint" | 6 of 6 | held |
| No caption asks for a photograph | none | none | held |
| The negations are restated, not kept | at most 2 of 6 captions keep "signature" | 1 (the Salt-Father's) | held |
| No caption asks for rendered text | no quoted span | none | held |
| Greedy decoding replays | the counted Reef-Runner caption equals the dry run's byte for byte | the same checksum | held |
| C′'s seconds | 5.28 to 6.46 | 5.94 | held |
| D′'s seconds | 2.32 to 3.00 | 2.71 | held |
| The caption costs more per picture | E's median over C′'s 1.03 to 1.30 | 1.20 | held |
| The shift costs nothing | F's median over E's 0.95 to 1.05 | 1.00 | held |
| Every caption within 45 seconds, the first one's load included | at most 45 | 24.49 at most | held |
| PUP-WIN | E above C′ on at least 4 of 6 | 5 | held |
| E above FLUX.1 [dev] on at least 4 of 6 (reported, never a rule) | at least 4 | 4 | held |
| The 2026-10-15 split re-reads at a fresh seed (tests little: a judge ranking at random lands here 50 times in 64) | C′ above FLUX.1 [dev] on 2, 3 or 4 of 6 | 2 | held |
| The judge can tell | names the upsampled picture on at least 4 of 6 | 4 | held |
| E shows its subjects | every "what must show" item ticked on all six E pictures | 10 of 10 | held |
| The newer shift moves a picture more than int8 did | median F-against-E gap at least 7.634 brightness steps (int8's largest on 2026-10-15) | 10.1 brightness steps | held |
What this does and does not say
- It says that on one NVIDIA GeForce RTX 3090 (24 GB) at 300 W, at 768 × 768, in the cove's style, on the game's six real subjects, one judge ranking blind at a fresh seed put FLUX 3 Image Klein reading its maker's upsampler captions above the same model reading our prompts as typed on 5 of 6, under a rule and odds written down first; and, reported, above FLUX.1 [dev] fp8 on 4 of 6 and FLUX.2 [klein] 4B on 5 of 6.
- It does not say how the maker's full current recipe ranks (E ran at α 6.93 and 768 × 768, the new workflows at α 16 and 4 megapixels; F adds the shift, not the size, and was not ranked), nor anything for another upsampler or caption source, more than one judge, one seed on the sheet, six subjects, one style or one card, or the game's own 16-bit FLUX.1 [dev] (A is fp8).
- It is not blind to the question. The judge was blind to the labels, but knew the run tested the upsampler and had seen FLUX 3's look on these prompts; the guess reports how often that told.
- It is not a verdict on the captions, which the judge never saw while ranking; they are printed for the reader to judge.
- It changes nothing for the game. A PUP-WIN buys this section, never a place: FLUX.1 [dev] stays the painter, and the line closed on 2026-10-15 stays closed unless the team reopens it with a new, separately registered test. FLUX 3 Image Klein ran here as an evaluation, which we read as non-commercial use under its licence; nothing it made feeds the game, the print lab or any model.
Sources for this section
No data kit ships, as above (the run's files name our machines); the settings, rule, captions, ranks and timings here are what a repeat needs. Software this section stood on, beside the list above: 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).
- [9] BFL, FLUX 3 Image Klein Prompt Upsampler model card and system prompts. https://huggingface.co/black-forest-labs/flux-3-image-klein-pup
- [10] BFL, the text-to-image workflow with the prompt upsampler: the file workflows/flux3_klein_bf16_t2i.json at revision feef60b0 (2026-10-17) of the FLUX 3 Image Klein for ComfyUI repository, whose card is [3]; the repository asks for contact details before it shows its files. https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy
- [11] BFL, FLUX 3 Image Klein reference code and README. https://github.com/black-forest-labs/flux-3-image-klein