Source: raw/GPT_Image_2.5_Tutorial_-_How_to_Use_ChatGPT_Image_Generator_Like_a_PRO.md (https://www.youtube.com/watch?v=iSkLZIGjo1c), YouTube auto-captions fetched 2026-10-30.
Creator: AI Master (youtube.com/@iamAImaster)
Platform: YouTube
Published: not recorded in the raw file
Sponsors and self-promotion
The tests run inside the creator’s own paid platform, AI Master (aimaster.me), which he promotes in the video. A separate segment is sponsored by Design.com. Most per-test results were shown on screen and not spoken, so the transcript gives the method in detail but few scores.
The creator runs GPT Image 2 and GPT Image 2.5 (the “Sunburst” variant, through the API) on the same prompts to check four of OpenAI’s claims: references hold, edits stay local, multi-turn edits keep their state, and complex prompts are followed. He then uses 2.5 alone on four client-style jobs. His conclusion: references hold better, edits stay more local and long prompts are easier to control, but text, product details and strict style rules still need a human check. This article records the test protocol and prompt habits, which are reusable for any image-model evaluation.
Key Takeaways
- OpenAI’s claims for 2.5 (as he relays them): it holds onto a person better when the scene changes; an edit touches only the region you name; it remembers image state across conversation turns; it follows complex visual instructions more closely; and latency can be “up to 50% lower” than GPT Image 2. OpenAI calls Sunburst “its most capable model for image generation and editing.”
- The speed claim did not show up in his timing. “On this run, that speed advantage didn’t show up… I didn’t see a consistent speed up here.” One prompt, not a benchmark.
- His verdict. “References hold up better, edits stay more local, and longer prompts are easier to control.” If you work from portraits, product shots or repeated edits, “that’s enough reason to move to 2.5.”
- Four prompt habits he used throughout:
- for edits, say what changes and what must stay untouched;
- for text, give the exact words and where they belong;
- for lighting, give a direction, not mood words;
- for a series, keep the visual rules fixed and change only the scene or format.
- He did not prove the habits work. “None of this proves these techniques improve the output because I never tested the same prompt without them.”
- Still check by hand: small or dense text, faces, labels and product shapes across angles, long edit chains, and strict style rules (“no gradients”, a two-colour palette).
The Four Comparison Tests (2 vs 2.5)
- Reference drift. Same portrait, then force a three-quarter angle and completely different lighting at once (“those two demands separate a model that rebuilds a face from one that only copies it”).
- Score only the face: jaw width against eye spacing, hairline shape, skin at full size (“airbrushed pores fail the run”), and whether the new street light actually lands on the face.
- Run three times per model and compare faces at the same size: “Comparing faces at different sizes is how you talk yourself into a match that isn’t there.”
- Add a mood board beside the face reference. Fail if the board changes the jawline or the portrait kills the board’s palette.
- Local edit (blink test). Change only a cup’s printed text. Flip between original and edit and watch what moves outside the cup: the hand first, then background and lighting. The new text must follow the cup’s surface without the cup being redrawn.
- Six-turn edit chain. One product shot, six turns, one change per turn, no restarts. At each turn ask: did it do the thing, and is earlier work intact?
- Turn 3: does the charcoal background drift back to grey; is the rim light still there?
- Turn 4: the instruction names only the reflection, so check the contact shadow survives. If it disappears, rerun with the shadow protected in the prompt to tell a hard limit from a prompting problem.
- Turn 6 against turn 1: label rotation (easy to “quietly reflatten”), metal brush cap, identical bottle geometry.
- Prompt adherence on an overloaded brief. Counted objects (exactly three standing bottles, one fallen lower left, a leaf between bottles two and three), light from the upper right, two text blocks in set corners, “no other objects in frame.”
- Read text at rendered size, not zoomed. “48H steep” was the hard case: “packaging work is where one wrong character costs money.”
- Run three times; “an image that looks great and misses the brief is the expensive kind.”
Four Jobs on 2.5 Alone
| Job | Brief | What he checks |
|---|---|---|
| Conference poster | One plain portrait; subject left, waist up; right half empty for the headline “build faster” and a small September date | Face vs reference, a genuinely empty right half, both text lines letter by letter, rim light, hands, hairline; three runs to tell “the model or luck” |
| Ad set | One product photo into square feed, vertical story and wide banner; top third clear for copy | Label artwork and text in each, bottle shoulders and cap, shadow matching the upper-left light; count how many are usable unfixed |
| SaaS dashboard (no reference) | Left nav, four metric cards, one line chart, a five-row transactions table | Layout coherence, then nav labels, card titles and numbers; chart labels against values; consistent dates, plans and amounts |
| Flat vector slide illustration | Two named brand colours, hard edges, no gradients or soft shadows | Symmetry and shape logic, any gradient or soft shadow, any extra tone, consistent outlines across three runs |
- “One good poster out of three is a screenshot. Three out of three is something I’d put in front of a client.”
- On the dashboard: “A mockup that looks right until you read it is the dangerous one. That’s the version that reaches a client deck before anyone notices.”
Try It
- Reuse his face test when choosing an image model for a recurring character: new angle plus new lighting in one prompt, three runs, faces compared at equal size.
- Use the blink test on every “local” edit before you accept it.
- For product campaigns, protect anything the instruction does not mention (shadows, reflections, label orientation) by naming it as “must stay unchanged.”
- Build ad briefs with a reserved copy zone (“right half empty”, “top third clear”) and reject outputs that put anything there.
- Read all generated text at its final display size, and count props against the brief, before a client sees the file.
- If speed matters to you, time your own prompts; his one-run check did not reproduce OpenAI’s latency claim.
Design.com Segment (Sponsored)
- Enter a business name and keywords; it generates logo options (his test brand, “Ember Oak Studio”, came back with several options “close to commercially ready”).
- Type edits in plain language; once the logo is confirmed it builds a brand kit, and every template inherits the logo’s colours and fonts.
- The free tier runs the full generation and brand-kit preview; downloads need an upgrade. “Premium starts at 300 to $1,500 for a basic identity, with no source given.
Open Questions
- The per-test outcomes for GPT Image 2 against 2.5 (which model won each comparison) were shown on screen and not stated in the transcript.
- Whether the ChatGPT app serves Sunburst or Flare by default; he tested Sunburst through the API only.
- No API pricing for 2.5 is given.
Related
- ChatGPT Images 2 (GPT Image 2) — Launch Coverage — OpenAI’s Images 2.5 announcement (Flare and Sunburst, the 50% latency claim).
- ChatGPT Images 2.5 — Field Test (MattVidPro) — a second hands-on test; he also found edit consistency is the main gain.
- Nano Banana — Google’s competing image line, including the same creator’s Nano Banana 2.1 tests.
- MiniMax Design — Motion Graphics From One Agent Workspace — GPT Image 2.5 used for style frames inside a video pipeline.
- awesome-gpt-image-2 — prompt library to test against.
- Picking the Right Model — Building Evals for Model Selection — turning ad-hoc comparisons like these into a repeatable eval.