Two days with GPT Image 2, and the author's conclusion is not about image quality at all: what this model eats is the visual-asset prep work that usually costs the most hours. Information graphics come out with near-perfect text layout on a single draw, product call-outs are one image plus one prompt, and a 25-panel contact sheet can feed a video graph downstream.

Evidence note: this episode ships no workflow JSON and 说明.txt records no workflow or mirror link — this is a review of a closed model, not a graph walkthrough. Figures below are the author's spoken observations.

When this is useful

  • Information graphics and posters where text must be spelled right and laid out sensibly.
  • Product explanation: one product photo in, a column of annotated call-outs out.
  • Period-accurate photography — the author singles out millennium-era snapshots as a high point.
  • Producing a 25-panel storyboard grid to feed into an LTX 2.3 multi-image first/last-frame graph.

How this differs from the previous build

Compared with the open image models, the difference is where the work happens.

  • Text layout is treated as solved. The author puts layout accuracy near 99% on the infographic tests, with a single draw per image, and speculates the model may be typesetting before rendering.
  • Prompting is conversational. There are no magic tokens; the author says the prompts came out of ordinary dialogue with the model.
  • Grid generation is the bridge to video. A 25-panel grid of a character transforming becomes 25 keyframes for LTX 2.3, which is how the author gets a transformation without generating each frame.

Running it

  1. Run it wherever your subscription applies — the author works inside an online canvas and inside an online ComfyUI, and notes the other major tiers also expose it.
  2. Write the prompt conversationally; iterate by describing what to change rather than by adding keyword soup.
  3. For infographics, specify the subject, the labels and the layout intent in one prompt and accept the first draw — that is where the near-99% figure comes from.
  4. For product explanation, feed one product image and ask for annotated call-outs.
  5. For video work, ask for a grid, then crop and feed the panels into an LTX 2.3 first/last-frame graph.
  6. Watch the density dial: pull back when the model starts over-packing the frame.

Honest limits

  • Over-dense output is a real flaw — the author describes it as looking like an image passed through repeated image-to-image rounds.
  • Celebrity likeness degrades over time as moderation tightens; the author notes one likeness had already drifted by the time he generated it.
  • Free tiers lose the image-with-thinking mode; the full behaviour is not available to them.
  • The 25-panel video experiment is explicitly a work in progress — the author calls the graph still crude and the result imperfect.
  • This is a closed model: no weights, no graph, no parameters to inspect or reproduce.

Episode resources