Accelerando in Sixteen Slides: GPT Image 2.0 at High and Low Quality

Before you read

Spoilers ahead. This page walks through the entire arc of Accelerando, from Manfred's first scene to the family's last. If you would rather meet the story cold, read it first and come back.

About the source. Accelerando is copyright © Charles Stross, 2005. Stross has made the complete novel free to read online under a Creative Commons Attribution-NonCommercial-NoDerivs 2.5 license. If these images make you curious, read the real thing; it is far better than any summary of it.

About these images. The slides below are a non-commercial fan tribute, generated to test an image model and to share enthusiasm for the book. They are not for sale, and they are no substitute for the text. The license above does not extend to illustrations, so I rely on what I believe is fair use; if the rights holder disagrees, I will take them down.

The Source Material

Accelerando (2005) by Charles Stross is a fix-up novel assembled from nine linked stories. It follows the Macx family across roughly fifty years and three generations as the rate of technological change tips from rapid to incomprehensible. The first part belongs to Manfred Macx, a “venture altruist” in near-future Amsterdam who gives away patentable ideas and lives entirely inside an augmented-reality cloud of agents. The second follows his daughter Amber, who escapes to Jovian space, rules a habitat called the Ring Imperium, and uploads her mind onto a starship the size of a soda can. The third lands on Amber’s son Sirhan, in a solar system being dismantled, atom by atom, into computing substrate by the family’s own posthuman descendants.

It is a useful test case for image generation precisely because it is hard to illustrate. The book is dense with abstractions (uploaded lobsters, Economics 2.0, the “Vile Offspring”) that have no obvious visual form. A summarizer has to commit to concrete pictures for ideas the author left deliberately slippery.

The Experiment

The two decks come from the same model. The only thing that changed between them is one setting.

I generated all sixteen slides with my own image-generation tool, driving OpenAI’s GPT Image 2.0 model (gpt-image-2) from a fixed prompt for each scene. Then I ran the entire deck a second time, changing exactly one variable: the quality parameter, from high to low. Everything else stayed identical: the prompt, the editorial layout instructions, the requested 3:2 aspect ratio.

The reason this comparison is worth running at all is economic. These APIs price image generation in tiers, and the high tier costs meaningfully more per image and takes longer to return. For a one-off hero image, nobody cares. For a tool that generates decks of a dozen or more slides on demand, the multiplier is the whole budget. So the practical question is narrow and concrete: when you pay for “high,” what do you actually get back?

One caveat before the pictures. Because each tier is a separate generation from the same prompt, not the same image rendered at two resolutions, some of the differences you will see are not quality at all. They are simply two different samples from a model that never draws the same scene twice. I have tried to separate “this tier is better here” from “these two happen to be framed differently,” but the line is not always clean. Where I am guessing, I say so.

In the comparisons below, the high quality slide is on the left and the low quality slide is on the right. Drag the divider to wipe between the two, or click any pair to open it full screen for a closer look.

Part One: Slow Takeoff (Manfred)

The opening slides establish Manfred in Amsterdam, his world thick with floating interface panels. This is the most demanding test of embedded text in the entire deck.

Low-quality slide: Making Strangers Rich — Making Strangers Rich. The high version packs denser, more atmospheric data panels; the low version's panels are sparser but noticeably more legible (NEARBY PERSONS, MARKET PULSE, CANAL LEVEL).

High-quality slide: Making Strangers Rich — Making Strangers Rich. The high version packs denser, more atmospheric data panels; the low version's panels are sparser but noticeably more legible (NEARBY PERSONS, MARKET PULSE, CANAL LEVEL).

The high slide renders more interface clutter, which reads as richer at a glance, but most of its microtext dissolves into plausible-looking gibberish. The low slide draws fewer panels and labels them with words you can actually read. If the goal is a believable augmented-reality overlay, “more” is not obviously “better.”

Low-quality slide: Take Us to the Ocean — Take Us to the Ocean. The uploaded lobsters phone Manfred to defect. High renders a darker, denser constellation of crustaceans; low is brighter and slightly sparser. Both are legible.

High-quality slide: Take Us to the Ocean — Take Us to the Ocean. The uploaded lobsters phone Manfred to defect. High renders a darker, denser constellation of crustaceans; low is brighter and slightly sparser. Both are legible.

Low-quality slide: Dismantle the Moon — Dismantle the Moon. Manfred pitches computronium over beer at De Wildemann's. Low renders the joke on the beer list (16 PAGES of BEER) and the Dyson-sphere schematic more cleanly than high.

High-quality slide: Dismantle the Moon — Dismantle the Moon. Manfred pitches computronium over beer at De Wildemann's. Low renders the joke on the beer list (16 PAGES of BEER) and the Dyson-sphere schematic more cleanly than high.

Low-quality slide: Twelve Million, Owed — Twelve Million, Owed. Pamela, Manfred's ex-fiancee and an agent of the IRS, presents the tax bill. Both render the $12,000,000.51 statement; low's version is the more readable document.

High-quality slide: Twelve Million, Owed — Twelve Million, Owed. Pamela, Manfred's ex-fiancee and an agent of the IRS, presents the tax bill. Both render the $12,000,000.51 statement; low's version is the more readable document.

Low-quality slide: The Suitcase Full of Noise — The Suitcase Full of Noise. Near-identical staging; the high version's particle field is denser and the lighting slightly moodier.

High-quality slide: The Suitcase Full of Noise — The Suitcase Full of Noise. Near-identical staging; the high version's particle field is denser and the lighting slightly moodier.

Low-quality slide: The Cat Keeps Watch — The Cat Keeps Watch. Aineko, the robot cat. Here high earns its name: the cat's cybernetic body is rendered explicitly, where low draws something closer to an ordinary tabby with glowing eyes.

High-quality slide: The Cat Keeps Watch — The Cat Keeps Watch. Aineko, the robot cat. Here high earns its name: the cat's cybernetic body is rendered explicitly, where low draws something closer to an ordinary tabby with glowing eyes.

Low-quality slide: The Baton Passes — The Baton Passes. Amber is born into Pamela's custody, a golden thread linking the generations. The two tiers are almost indistinguishable in mood and composition.

High-quality slide: The Baton Passes — The Baton Passes. Amber is born into Pamela's custody, a golden thread linking the generations. The two tiers are almost indistinguishable in mood and composition.

Slide six is the clearest case in the deck where the high tier does something the low tier does not: it takes the word “robot” in “robot cat” literally and renders the machinery. That is a real difference in how much detail the model commits to, not a sampling accident.

Part Two: Point of Inflection (Amber)

The story now leaves Earth, and the slides shift from cluttered interiors to vast machines and deep space. This is where the high tier’s appetite for detail has the most room to work, and where the two tiers most often diverge in framing rather than fidelity.

Low-quality slide: It's All Mine, I Rule — It's All Mine, I Rule. Teenage Amber claims her first asteroid aboard the orphanage ship Ernst Sanger. Both render the ASTEROID LOCK panel; low's reads more cleanly and personalizes it (AMBER S.).

High-quality slide: It's All Mine, I Rule — It's All Mine, I Rule. Teenage Amber claims her first asteroid aboard the orphanage ship Ernst Sanger. Both render the ASTEROID LOCK panel; low's reads more cleanly and personalizes it (AMBER S.).

Low-quality slide: Queen of the Ring Imperium — Queen of the Ring Imperium. A clear divergence in interpretation: high draws a face-on, snowflake-symmetric habitat wreathed in plasma; low draws a three-quarter industrial view. Both are striking; neither is obviously the better rendering.

High-quality slide: Queen of the Ring Imperium — Queen of the Ring Imperium. A clear divergence in interpretation: high draws a face-on, snowflake-symmetric habitat wreathed in plasma; low draws a three-quarter industrial view. Both are striking; neither is obviously the better rendering.

Low-quality slide: A Bar That Doesn't Exist — A Bar That Doesn't Exist. Amber's uploaded crew drink in a simulated bar as the Field Circus accelerates toward an alien router. Low pushes the relativistic starbow harder into red; high keeps the bar warmer.

High-quality slide: A Bar That Doesn't Exist — A Bar That Doesn't Exist. Amber's uploaded crew drink in a simulated bar as the Field Circus accelerates toward an alien router. Low pushes the relativistic starbow harder into red; high keeps the bar warmer.

Low-quality slide: What the Router Holds — What the Router Holds. The crew meet the Wunch, parasitic uploaded brokers. Both render the same intricate red-and-cyan node map and a lone human silhouette for scale. The high version is marginally crisper in the fine wireframe.

High-quality slide: What the Router Holds — What the Router Holds. The crew meet the Wunch, parasitic uploaded brokers. Both render the same intricate red-and-cyan node map and a lone human silhouette for scale. The high version is marginally crisper in the fine wireframe.

Low-quality slide: Homecoming to Computronium — Homecoming to Computronium. The inner planets, disassembled. This pair is the closest match in the deck: same Dyson swarms, same orange galactic core, same diffraction sparkle.

High-quality slide: Homecoming to Computronium — Homecoming to Computronium. The inner planets, disassembled. This pair is the closest match in the deck: same Dyson swarms, same orange galactic core, same diffraction sparkle.

Low-quality slide: The Curator on the Lily-Pad — The Curator on the Lily-Pad. Sirhan watches his family arrive at Saturn. High frames the text in a black side-panel; low runs the image full-bleed and lets the type sit over it. A layout choice, not a quality gap.

High-quality slide: The Curator on the Lily-Pad — The Curator on the Lily-Pad. Sirhan watches his family arrive at Saturn. High frames the text in a black side-panel; low runs the image full-bleed and lets the type sit over it. A layout choice, not a quality gap.

Slides nine and thirteen are where this exercise gets slippery. The differences are large and immediately visible, but they are differences in composition (where the camera sits, how the page is laid out), not in quality. Two draws from the same prompt simply went different ways. If you generated the high tier twice, you would see this much variation between the two high images alone.

Part Three: Singularity (Sirhan)

The final act is the hardest to picture: an alien con-artist, the spreading hunger of Economics 2.0, and the family’s reckoning with the cat that has been managing them all along.

Low-quality slide: The Slug and the Appetite — The Slug and the Appetite. Manfred and Pamela, resurrected, face something vast and indifferent. Low resolves the looming face more clearly and labels the environment (ECONOMICS 2.0, GROWTH OPTIMIZATION); high leaves it more abstract.

High-quality slide: The Slug and the Appetite — The Slug and the Appetite. Manfred and Pamela, resurrected, face something vast and indifferent. Low resolves the looming face more clearly and labels the environment (ECONOMICS 2.0, GROWTH OPTIMIZATION); high leaves it more abstract.

Low-quality slide: Without Adult Supervision — Without Adult Supervision. The closing reconciliation, with Aineko and a coffee mug as witnesses. Both render the long italic quotation cleanly. Low adds more readable corner telemetry (SATURN, LIFE SIGNS, COFFEE WARM).

High-quality slide: Without Adult Supervision — Without Adult Supervision. The closing reconciliation, with Aineko and a coffee mug as witnesses. Both render the long italic quotation cleanly. Low adds more readable corner telemetry (SATURN, LIFE SIGNS, COFFEE WARM).

The last slide is a fitting place to end, because it isolates the one thing both tiers do reliably well: a long, italic, multi-line block of quoted text, set as display type, rendered without a single garbled word.

Observations

Long-form text rendering is, for practical purposes, solved at both tiers. Every headline and every paragraph of body copy across all sixteen slides is clean and legible in both high and low. This is the largest change from the previous generation of these models. In an earlier comparison using gpt-image-1.5, OpenAI’s output was riddled with artifacts like “encllhdfud” and “pananted” that broke immersion on text-heavy slides. That failure mode is essentially gone here. For editorial layouts with real captions, GPT Image 2.0 is a different class of tool.

Where the tiers diverge, it is in dense diegetic microtext, and low often wins. The interface panels, tax statements, asteroid-lock readouts, and HUD telemetry are the remaining frontier of unreliable text. The high tier tends to draw more of this microtext and garble more of it; the low tier draws less and keeps it legible. If your scene depends on believable embedded UI, the cheaper setting was frequently the better one in this deck. I would not assume that generalizes to every prompt, but it held consistently across these sixteen.

The high tier’s real advantage is committed detail. The clearest case is the robot cat (slide six), where high rendered the cybernetic body and low did not. High also holds up marginally better on the finest wireframe detail (slide eleven). When a prompt hinges on a small, literal, mechanical element, the high tier is likelier to honor it.

Much of the visible difference is sampling, not quality. Because each tier is an independent generation, the most dramatic contrasts (the face-on versus three-quarter Ring Imperium, the panel versus full-bleed layout on the Lily-Pad) are framing choices, not fidelity. It is easy to mistake “this one happened to compose better” for “this tier is better.” Over a large enough deck the two effects average out; over a single image they are hopelessly entangled.

Bottom line. For this kind of work (narrative editorial slides with headlines, body text, and atmospheric science-fiction scenes), the low quality setting of GPT Image 2.0 produced output that is difficult to distinguish from high, and occasionally better, at a fraction of the cost and latency. The high tier earns its premium only in two narrow situations: when a prompt depends on a specific literal detail being rendered, and when you can afford to generate several samples and keep the best composition. If you are generating decks at volume, the rational default is low, with selective re-rolls at high for the handful of slides that need them. I went in expecting the opposite.