{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "fifty-seven-milliseconds",
 "title": "Fifty-Seven Milliseconds: What a Diffusion Step Actually Buys",
 "dek": "One picture, drawn rung by rung up the step dial, on two painters that could not be more different. Timed to the millisecond and measured to the pixel, twice, a month apart — and the two runs say two different things, which turns out to be the finding.",
 "published": "2026-08-25",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "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/fifty-seven-milliseconds/",
  "md_url": "https://research.strata2signal.com/fifty-seven-milliseconds/index.md",
  "md_sha": "e58812fe28b0bed9ca6f0e5f6ccc1728e01e62e45df8f23bb2eda9a877602eeb",
  "html_sha": "1a395b892654bf808ae369abce7d29fb4e6f264bd6257ca09cbfcb776faff97f",
  "anchors_sha": "965f5e4b208d7614c2fd776166812425d7f4415b108cefdffc614c52818b42e8"
 },
 "short": {
  "paragraph": "One picture, drawn rung by rung up the step dial, on two painters that could not be more different. Timed to the millisecond and measured to the pixel, twice, a month apart - and the two runs say two different things, which turns out to be the finding.",
  "from_dek": true,
  "counts": {
   "words": 2839,
   "minutes": 13,
   "tables": 1,
   "kit": false
  },
  "bullets": [
   {
    "text": "the 20→24 rung moves the picture about 1.6×",
    "figure": "1.6×",
    "cite": "finding-two-the-first-turn-of-the-dial-is-the-big-one-and",
    "quote": "On the shipped painter (the 512² run), the 20→24 rung moves the picture about 1.6× as much as the typical later rung (a median change of 14.7 on the 0-255 scale, against 8.6 to 9.9 for the rungs after).",
    "span": [
     6702,
     6707
    ]
   },
   {
    "text": "the distilled painter is the fastest cell at 0.62 s",
    "figure": "0.62 s",
    "cite": "finding-four-the-argument-with-our-own-settings-and-what-it",
    "quote": "One more receipt, and it cuts against us: the distilled painter is the sharpest and fastest cell in the whole second run (0.62 s, finished, at 4 steps) and it did not take the tier - it paints a legible cursive signature on two of the six subjects - an inland figure and a foe - in every step column, because its template zeroes the negative conditioning that the no-signature guard rides on.",
    "span": [
     12334,
     12340
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "what-a-step-is-in-one-paragraph",
   "heading": "What a step is, in one paragraph",
   "level": 2,
   "span": [
    2379,
    3066
   ],
   "chunks": [
    [
     2379,
     3066
    ]
   ],
   "chars": 687,
   "digest": "a diffusion model un-ruins a picture by removing noise in a loop. each pass is a step, and for a fixed sampler and configuration, time should be a straight line with a slope. this is true for both painters.",
   "digest_skipped": null
  },
  {
   "id": "finding-one-a-step-is-a-fixed-price-good",
   "heading": "Finding one: a step is a fixed-price good",
   "level": 2,
   "span": [
    3066,
    5760
   ],
   "chunks": [
    [
     3066,
     5760
    ]
   ],
   "chars": 2694,
   "digest": "a step costs 56.9 milliseconds at 512² on the 24G VRAM rig. the slope holds whether fitting medians, all forty-two renders, or the post-exclusion set. the distilled painter, FLUX.2 [klein] 4B, costs 73 milliseconds a step over a quarter-second floor. the 30-step rung is slower than the 32-step rung.",
   "digest_skipped": null
  },
  {
   "id": "finding-two-the-first-turn-of-the-dial-is-the-big-one-and",
   "heading": "Finding two: the first turn of the dial is the big one — and how big depends on the painter",
   "level": 2,
   "span": [
    5760,
    8170
   ],
   "chunks": [
    [
     5760,
     8170
    ]
   ],
   "chars": 2410,
   "digest": "the mean absolute per-pixel difference measures how much a picture changes. on the shipped painter, the 20→24 rung moves the picture about 1.6× as much as a typical later rung. on the distilled painter, the 4→8 rung moves the picture 2.6× a later rung.",
   "digest_skipped": null
  },
  {
   "id": "finding-three-the-ladder-settles-on-one-painter-and-never",
   "heading": "Finding three: the ladder settles on one painter and never on the other",
   "level": 2,
   "span": [
    8170,
    10498
   ],
   "chunks": [
    [
     8170,
     10498
    ]
   ],
   "chars": 2328,
   "digest": "on the distilled painter, every one of the six subjects walks monotonically toward its final image. on the shipped painter, not one of the six subjects approaches its 40-step image monotonically. below a painter’s finishing point, the picture converges, but above it, each rung buys a sibling.",
   "digest_skipped": null
  },
  {
   "id": "finding-four-the-argument-with-our-own-settings-and-what-it",
   "heading": "Finding four: the argument with our own settings — and what it changed",
   "level": 2,
   "span": [
    10498,
    13022
   ],
   "chunks": [
    [
     10498,
     13022
    ]
   ],
   "chars": 2524,
   "digest": "the engine’s shipped sketch tier moved to 768² at 12 steps, which is 1.51 s against 3.48 s at the old setting. the distilled painter is the sharpest and fastest cell in the second run, at 0.62 s, finished, at 4 steps.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    13022,
    13979
   ],
   "chunks": [
    [
     13022,
     13979
    ]
   ],
   "chars": 957,
   "digest": "a denoising step is a fixed-price good, such as 56.9 ms. the biggest change is the first turn up from the floor, which is 1.6× a later rung on a conventional painter and 2.6× on a distilled one. find where a painter crosses from convergence into siblings and stop there.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work-and-see-it-live",
   "heading": "How to check our work — and see it live",
   "level": 2,
   "span": [
    13979,
    17231
   ],
   "chunks": [
    [
     13979,
     17231
    ]
   ],
   "chars": 3252,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    17231,
    18390
   ],
   "chunks": [
    [
     17231,
     18390
    ]
   ],
   "chars": 1159,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-d0af48de",
   "q": "how does a diffusion model create an image?",
   "a": "A diffusion model un-ruins an image by starting from pure noise and iteratively removing noise to match a prompt. Each iteration in this loop is called a step, and each step performs a consistent amount of work.",
   "cites": [
    "what-a-step-is-in-one-paragraph"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-4c950084",
   "q": "how much does the first turn of the dial change a picture?",
   "a": "The first turn of the dial provides the largest visual change. On a conventional painter, the 20→24 rung moves the picture about 1.6× as much as a typical later rung. On a distilled painter, the 4→8 rung moves the picture 2.6× a later rung.",
   "cites": [
    "finding-two-the-first-turn-of-the-dial-is-the-big-one-and"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-1195adea",
   "q": "how do different painters behave as step counts increase?",
   "a": "On the distilled painter, every subject walks monotonically toward its final image as steps increase. On the shipped painter, not one of the six subjects approaches its 40-step image monotonically, with every subject stepping away from the target at least once.",
   "cites": [
    "finding-three-the-ladder-settles-on-one-painter-and-never"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-7dde6cf1",
   "q": "how did research change the sketch tier settings?",
   "a": "The sketch tier moved to 768² at 12 steps. This change provided more picture and less time compared to the old setting of 512² at 32 steps, as the jump from 8 to 12 steps is where the picture arrives.",
   "cites": [
    "finding-four-the-argument-with-our-own-settings-and-what-it"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-277304e8",
   "q": "what are the key takeaways regarding denoising steps?",
   "a": "A denoising step is a fixed-price good. The biggest change occurs during the first turn of the dial. Below a painter's finishing point, the picture converges toward its final form, but above that point, rungs produce siblings rather than improvements.",
   "cites": [
    "what-to-take-with-you"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "a-rig-your-friend-already-owns",
   "why": "shares ground with § Finding four: the argument with our own settings — and what it changed · § The sketch ladder — the number the game lives on"
  },
  {
   "slug": "how-a-vision-model-sees",
   "why": "shares ground with § Finding four: the argument with our own settings — and what it changed · § Finding four: the render payoff — and what the bake-off measures"
  },
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § Finding four: the argument with our own settings — and what it changed · § Renders"
  }
 ],
 "thanks": "",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 2,
  "digests_written": 6,
  "digests_kept": 6,
  "chips_written": 6,
  "chips_kept": 5,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 2,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0
  },
  "new_noun_tokens_checked": [
   "0.62",
   "1.51",
   "1.6×",
   "12",
   "2.6×",
   "20",
   "24",
   "24G",
   "3.48",
   "30-step",
   "32",
   "32-step",
   "4",
   "40-step",
   "4B",
   "512",
   "56.9",
   "73",
   "768",
   "8",
   "96G",
   "99",
   "FLUX.2",
   "VRAM"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "a step costs 56.9 ms",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "chip",
    "reason": "figure",
    "item": "A step costs 56.9 milliseconds at 512² on the 24G VRAM rig. For a distilled painter at 768² on the 96G VRAM workstation, a step costs 73 milliseconds, while the shipped painter on the same canvas cost",
    "detail": "figures absent from cited spans: ['768', '96']"
   }
  ]
 },
 "generated_utc": "2026-09-09T01:27:17Z"
}