Jonathon's AI Wiki

ChatGPT Images 2.5 — Field Test (MattVidPro)

6 min read

Source: raw/I_Gave_ChatGPT_Images_2.5_Some_VERY_Weird_Prompts.md. Creator: MattVidPro AI. URL: https://www.youtube.com/watch?v=Q5-LJDxZlrg. Platform: YouTube (auto-captions, fetched 2026-09-29; upload date not recorded). It mixes his own tests with images from his Discord community.

MattVidPro’s hands-on test of OpenAI’s Images 2.5 in ChatGPT, covering stop-motion sequences, web-referenced likeness, positional-comment editing, flowcharts and a head-to-head with Google’s next Nano Banana. His verdict is “definitely an upgrade” but not a generational leap: “I understand why they wouldn’t call this an images 3.0.” This article covers only the field test. For what OpenAI announced for 2.5, see the Images 2.5 section of ChatGPT Images 2 (GPT Image 2) — Launch Coverage.

Key Takeaways

  • Edit consistency is the real gain. Next to a GPT Image 2 sequence that is “bouncing all over the place,” “2.5 is dramatically smoother.” That makes frame-to-frame stop-motion workable; community examples included a claymation dragon hatching and Will Smith eating spaghetti.
  • The first image is the sharpest. “That first image is going to give you your sharpest, most detailed quality. Then any second image is going to have a little bit of [degradation].” The biggest drop is between the first and second image; after that the model “is good at maintaining itself … you won’t see it get much worse than this.”
  • It pulls reference images from the web on its own. It is “much better at grabbing references from the web and actually using them automatically”: asked for “Matt Vidpro doing whatever,” it “go[es] grab images of my face.” When he told it not to search the web, it produced only “a loose description of me.”
  • Positional comments help, but targeted fixes take rounds. He pinned comments to parts of the image to fix an impossible third swing cable. It took several prompts, the final cable was attached in the wrong place, and the image was “still degrading over time.”
  • Sketch input works. A loose sketch of river, house, grass and mountains became “a fully coherent and quite gorgeous oil painting” (a Discord member’s example).
  • Flowcharts: simple yes, complex no. A pumpkin-growing flowchart was “surprisingly good” and correct by his check. A complex branching chart reversed its yes/no labels and had steps that “trap you in an infinite loop.” His read: “I suspect 2.5 isn’t that much better” than 2.0 on complex charts.
  • Old artefacts remain. The “hallucinatory pattern issue … seems to be just as prevalent as before,” and “the yellow tinged filter also appears to still be in the images.” He guesses the pattern marks images as AI-generated; that is his speculation.
  • Text and UI rendering are strong. A retro computer-class scene had legible cups, posters, books and board. A RuneScape-style screenshot had UI “kind of nearly perfect.” A Twitch screenshot came out “mostly close to perfect,” including on-topic chat comments.
  • Competitor note. Google’s next Nano Banana, codenamed “Spicy Mayo,” “definitely beats Nano Banana 2” on his Minecraft-screenshot test. Images 2.5 did “a lot better than I thought it would” on the same prompt, and he concludes “these two models could actually be closer than we think.”

Details

  • Stop-motion through Codex.
    • Following a community member’s method, he asked Codex to generate the frames and assemble a Lego stop-motion clip: him on an Indiana Jones-style run from a rolling lemon. “37 minutes later and we get our result.”
    • Output was “very coherent still” but “the whole image is getting crunched,” with “some pretty big inconsistencies.”
    • Caveat: “Even Codex itself couldn’t confirm whether this was images 2.5.”
  • Speed. He says 2.5 is faster than 2.0 but gives no timing.
  • Community realism (images by Discord mod “Rorow Tuck”, caption spelling).
    • Convincing phone-snapshot style, with errors in anatomy (a twisted back) and grip.
    • A quirk: the model is “kind of obsessed with toilets.”
    • He notes “Images 2.0 was very similar” for this style, so realism is not where 2.5 differs.
  • Likeness from an upload. An uploaded photo turned into a “lemon villain” “used my likeness from the provided picture very well.” That same image was where the hallucinatory pattern was “visible from afar.”
  • Who notices the difference. “A lot of casual users might not really notice a big difference.” Heavy users on his Discord “continuously seem to think that this is a pretty big upgrade.” He attributes the gain to “the intelligence under the hood and understanding those highly complicated prompts.”

Try It

  1. Make the first generation count. Put the full brief in the first prompt and treat that output as the hero asset, since it has the most detail. Expect one visible quality drop on the first edit, then little further loss.
  2. Check likeness with and without web search. For a brand, product or named person, run the same prompt both ways. With search on, it pulls public images of a named person automatically; decide whether that is acceptable before using it for client work
  3. Use positional comments for local fixes, and budget several rounds. One comment per problem region; check structural details such as cables, symmetry and straight lines after each pass.
  4. Keep flowcharts simple. Check every yes/no branch by hand before publishing a generated flowchart. For complex logic, build the chart in a diagram tool
  5. Try stop-motion. Ask Codex to generate a frame sequence and stitch it into a video, and expect about half an hour for a short clip. Confirm which image model produced the frames.
  6. Plan a colour pass if the warm, yellow cast doesn’t suit your brand palette ^[inferred]. He thinks OpenAI simply “prefers warmer images.”

Open Questions

  • Does Codex’s image-generation path use Images 2.5 or 2.0? The source couldn’t confirm.
  • Is the recurring “hallucinatory pattern” a deliberate AI-image marker? That is MattVidPro’s guess only.
  • “Spicy Mayo” is a codename. Its release name, date and pricing are not in the source.
  • The source gives no API availability, pricing or latency figures for Images 2.5. See the launch article’s Images 2.5 section for OpenAI’s own claims.

Graph View

Backlinks