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)
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.
2,776 words, about 13 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 is the 768 × 768 picture as the model drew it, saved as a JPEG; select one to open it at full size. 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.





















































