The bench
FLUX 3 Image Klein, run again with its maker's prompt upsampler
exhibit ninety-five The bench
Published 2026-10-27 (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 and, from about 22:58Z that day, says not to use it without its prompt upsampler, a second model that rewrites a short prompt into a long caption. Our blind bake-off that day ran with the upsampler off. On 2026-10-24 we ran the model again with the upsampler writing the caption, on the same six subjects from our game, and the same judge, ranking blind at a fresh seed (a seed fixes the starting noise), put the caption's picture above the as-typed picture on 5 of the 6; the rule we wrote down first needed 4. FLUX.1 [dev] stays the game's painter. This page is the short form of a dated section on the full exhibit, linked below.
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
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.
- The plan was frozen at 04:14:23Z (see The rule, written down first)
- The counted run took 19 minutes 31 seconds (see How it ran)
2,491 words, about 11 minutes to read.
The summary is this page’s own; the receipt lines were drafted by a model on this workshop’s network and every figure in them is in the article, checked before this page went out — 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. This page distils one dated section of FLUX 3 Image Klein, ranked blind in a game's own painted style, the full exhibit, which holds the original 2026-10-15 run, every table and every picture we may show (FLUX.1 [dev]'s are held); every figure here is printed there first.
What changed, and why a caption matters
RealKeep, our small role-playing game, paints its world in one house style: stylized oil-painted realism, loose quick brushwork, in the manner of Dishonored game art (a 2012 video game, named only for its style). On 2026-10-15, FLUX 3 Image Klein's release day, we ran it beside FLUX.1 [dev], the game's painter, and FLUX.2 [klein] 4B; ranking blind, the judge put it above FLUX.1 [dev] on 3 of 6 subjects, one short of the bar. That run turned the upsampler off by our registered rule, so that all four models read the same words.
The maker says the upsampler is not optional. 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). On release day BFL's reference code and its own workflow for ComfyUI, a free picture-making program, ran it by default, and its model card said a short prompt gives generic output. 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". So the 2026-10-15 ranks stand for prompts as typed and may understate FLUX 3.
On our six prompts the upsampler turns a one-line subject into a caption of 228 to 266 words (median 244), or 293 to 357 tokens (a token is a word or part of one). All six keep "Dishonored", "oil-painted" and "brushwork", and none asks for a photograph or quotes text to draw. The caption costs about a fifth more per picture (a median 7.14 against 5.94 seconds, seeds 202 and 303). One is printed in full below; all six are on the full exhibit.
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.
- Two arms on the rule. C′: FLUX 3 Image Klein, distilled, BF16, reading our prompt as typed (the game's style clause, a comma, the subject's clause), the same files and ComfyUI graph as the first run. 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. Beside them on the sheet, reported and deciding nothing: FLUX.1 [dev] in its 8-bit floating-point (fp8) build (32 steps, guidance 3.5, the game's recipe) and FLUX.2 [klein] 4B (BF16, 8 steps).
- The caption. The upsampler ran alone on the card first, with BFL's own workflow settings: its text-to-image system prompt (standing instructions read before every request; 3,404 characters), greedy decoding (always the top-scoring next token, so a request gets the same caption at every seed) and thinking off. The six captions were checksummed before the first E picture; none was re-rolled or edited.
- The sheet. One per subject: the four arms' seed-404 pictures, shuffled, unlabelled, resampled to 768 pixels, metadata stripped, the key sealed and its checksum printed on every sheet; it shows the typed prompt, never a caption, and the judge ranks its four pictures from 1 (best) to 4 by which best reads as the game's style while showing the subject. Seed 404 is fresh: the judge, the member of our team who directs the game's art, had already seen the first run's pictures at seeds 101, 202 and 303 labelled, FLUX.1 [dev]'s included.
- 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. In the frozen plan's words, "A PUP-WIN buys a sentence on the page, never a place in the game or the print lab", and the line of testing closed on 2026-10-15 stays closed.
- 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.
How it ran
It ran on one NVIDIA GeForce RTX 3090 (24 GB), the same board as the first run, alone on the card, its power limit read at 300 W before every arm (highest readings 69 °C and 291 W); ComfyUI v0.42.5 (v0.41.0 on 2026-10-15), PyTorch 2.13.0, offline, every model loading whole. 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. On the newer ComfyUI the 36 as-typed FLUX 3 pictures at seeds 101, 202 and 303 (in BF16 and in its maker's 8-bit integer build, int8) came out pixel for pixel the same as the first run's lossless originals (0.000 brightness steps, out of 255, apart on every one). 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.
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; no bar was set for either, and no ties were written.
Three of the pictures
The rule's pair on three subjects: two where the caption's picture placed first and the as-typed picture last (Wrenna Sault) or second (the Salt-Father), and the one where the as-typed picture placed above the caption's (the breakwater). 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. 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); the captions under Apache 2.0. FLUX.1 [dev]'s pictures are held by a ruling of ours and appear in words only. The full exhibit shows, for all six subjects, the three seed-404 sheet pictures it may show (as typed, the caption and klein 4B's), with the caption under each table.
Portrait · Wrenna Sault, seed 404
| seed | C′ · FLUX 3 Image Klein, BF16, the prompt as typed 16 steps · median 5.94 s | E · the same, reading the upsampler's caption 16 steps · median 7.14 s |
|---|---|---|
| 404 |  5.96 s |  7.36 s |
The typed prompt, after the style clause, was "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"; the judge put the as-typed picture 4th of 4 and the caption's 1st.
The caption E read (246 words, the upsampler's output from the typed prompt): 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.
Foe · the Salt-Father, seed 404
| seed | C′ · FLUX 3 Image Klein, BF16, the prompt as typed 16 steps · median 5.94 s | E · the same, reading the upsampler's caption 16 steps · median 7.14 s |
|---|---|---|
| 404 |  5.89 s |  7.26 s |
The typed prompt asks for "no face" and, in the style clause, "no signature, no watermark". Counted by eye, the as-typed picture carries a scrawl like a signature at the bottom centre and the caption's does not; its caption is the one of six that keeps the words ("unmarked by typography, emblems, signatures, or watermarks"). Both pictures were ticked for "no face", as every "what must show" item was on all six as-typed and all six caption pictures. The judge put the caption's picture 1st and the as-typed 2nd.
Scene · the sea-fog breakwater, seed 404
| seed | C′ · FLUX 3 Image Klein, BF16, the prompt as typed 16 steps · median 5.94 s | E · the same, reading the upsampler's caption 16 steps · median 7.14 s |
|---|---|---|
| 404 |  5.98 s |  7.08 s |
The one subject of six where the as-typed picture ranked above the caption's: 1st and 2nd of 4, both above FLUX.1 [dev]. The judge's guess on this sheet fell on the as-typed picture.
What this cannot say
- It is one judge, one seed on the sheet, six subjects, one style, one card. Nothing here speaks for another upsampler or caption source, more judges, other seeds, styles or cards, or for the game's own 16-bit FLUX.1 [dev] (the sheet's is the fp8 build).
- It is not the maker's full current recipe. E ran at α 6.93 and 768 × 768; on 2026-10-17 BFL's new workflows moved the timestep shift (how many of the 16 steps are spent at high noise) to α 16, and the default size to 4 megapixels. An arm with that shift at our size was measured, never ranked: its pictures' pixels sit on average 6.5 to 14.7 brightness steps (out of 255) from E's at the same subject and seed (median 10.1, 24 pairs).
- 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 a sentence on the page, never a place: FLUX.1 [dev] stays the painter, and the line of testing 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.
The full exhibit
Everything above is a distillation. On FLUX 3 Image Klein, ranked blind in a game's own painted style: the original run of 2026-10-15, with its rules, ranks, seconds, memory and the 54 pictures FLUX 3 and klein 4B made; and the dated section of 2026-10-24, with every arm's seconds, the 18 seed-404 sheet pictures it may show (as typed, the caption and klein 4B, one table per subject), the six captions in full, and the 23 predictions frozen with the rule, of which 22 held and 1 was refuted. No data kit ships, as there: the run's files name our machines; the settings, rule, captions, ranks and timings printed there are what a repeat needs.
Who ran this, and thanks
Thanks 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.
Sources
Sources 1 to 4 are the full exhibit's, cited there as [2], [3], [9], [10] and [11]; this page adds no source of its own.
- 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, and its workflow file workflows/flux3_klein_bf16_t2i.json at revision feef60b0 (2026-10-17); the repository asks for contact details before it shows its files. https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy
- BFL, FLUX 3 Image Klein Prompt Upsampler model card and system prompts. https://huggingface.co/black-forest-labs/flux-3-image-klein-pup
- BFL, FLUX 3 Image Klein reference code and README. https://github.com/black-forest-labs/flux-3-image-klein
- The full exhibit, with its dated section of 2026-10-24. https://research.strata2signal.com/flux3-klein-in-the-coves-style/#upsampler-2026-10-24
- RealKeep. https://realkeep.strata2signal.com/
Corrections and later measurements will be added below, each dated (UTC), saying in plain words what it counts.