Source: raw/New_BEST_AI_image_generator_is_here.md. Creator: AI Search. URL: https://www.youtube.com/watch?v=q9tpIc8PVKM. Duration: about 32 min. Platform: YouTube (auto-captions, fetched 2026-10-30; uploaded 2026-09-10). Sponsored by Higgsfield. Caption fixes: “GPZ / GBT image” → GPT Image, “Cance / Miniax” → Seedance / MiniMax, “topography” → typography (where he means type design).
AI Search ran about 24 hard prompts through GPT Image 2.5 (the high-quality Sunburst variant), GPT Image 2 and Google’s Nano Banana 2, after demonstrating 2.5’s new sketch input, annotation edits and transparent layers assembled into a PSD. His verdict: 2.5 is “slightly better” than GPT Image 2 in most cases, “not a huge significant leap,” and both GPT models beat Nano Banana 2 on nearly every test. The useful findings for production work are the layer and transparency tricks, and the categories where 2.5 is worse than 2 (dense grids, chart conversion, UI redesign). All judgments are his, from single generations.
Key Takeaways
- Transparent layers are the practical new trick. 2.5 generated five sequential transparent layers (foreground to background) of a lake scene and assembled them into a PSD in about 4 minutes. It also decomposed an existing poster into separate transparent layers (background, person, each text element) in a PSD in about 6 minutes, so each element could be moved and resized.
- Draw or annotate instead of describing. ChatGPT’s new Sketch tool (plus menu → Sketch) turned a crude drawing into a drone photo of an ocean-view house; writing edit instructions directly on a photo and asking 2.5 to follow the annotations also worked.
- Multi-turn consistency is better: a cube rotated over 150 frames stayed consistent, which is what community stop-motion animations rely on.
- Where 2.5 won (his calls): YouTube homepage mockup, weekly outfit infographic, sprite sheet (slightly), exploded device diagram, floor-plan view (least wrong), Where’s Waldo (least wrong), clock-and-wine-glass, a product held by an influencer with small text preserved (but a malformed hand).
- Where GPT Image 2 won: a 100-poster anime grid (2.5’s faces were messed up), table-to-bar-chart conversion (2.5 misread one value), and a landing-page redesign (2.5 “decided to add way too much detail”).
- Style difference: 2.5 “tends to add a lot more details” you did not ask for and looks more natural and amateur; GPT Image 2 is more literal and polished. 2.5 drops the yellowish tinge he still sees in GPT Image 2.
- All three still fail Chinese/English street signage, a labelled biology worksheet, endemic Borneo frog species, a correct mate-in-two chess diagram, and a camera view from a specific door on a floor plan.
- Pricing he reads from the API page: about 0.5¢ for a low-quality 1024×1024 image, 21¢ at max quality, and roughly 40¢ for a max-quality 4K 16:9 image ^[ambiguous] (the caption drops the unit). Flare and Sunburst cost the same; Flare has about 50% lower latency.
Results by test
| Test | Winner (his call) | Note |
|---|---|---|
| Grid of 100 anime posters with names | GPT Image 2 | Nano Banana 2 had web search but produced gibberish posters |
| Messy Windows 11 desktop (Slack, Gmail, Excel, PowerPoint) | Tie, GPT 2.5 and 2 | Both Excel sheets added up correctly; Nano Banana 2 worst |
| YouTube homepage for a tech bro | GPT 2.5 | Narrow margin |
| Brand identity board for an eco matcha brand | Tie, all three | Logo grid, mood board, palette, type, mock-ups all present |
| 7-day women’s outfit infographic | GPT 2.5 | GPT 2 close |
| 5×5 sprite sheet, princess sprinting then slashing | GPT 2.5 (slightly) | Nano Banana 2’s grid could not be animated |
| Complex table to bar charts | GPT Image 2 | 2.5 misread one total-response-time value (the caption garbles the numbers); Nano Banana 2 mangled logos and legend |
| TikTok livestream screenshot | Tie, GPT models | 2.5 more casual, 2 more polished |
| Full typeface from a few sample words | Close, GPT models | 2.5 changed the W; 2 added “hello world” |
| Exploded, labelled phone diagram | GPT 2.5 | Accuracy not checked |
| Black-and-white manga fight page from two references | None good | 2.5 over-detailed; 2 more coherent but yellow |
| Landing-page redesign from a screenshot | GPT Image 2 | |
| Ad storyboard from a product image | Tie, GPT models | Feed it to a video model afterwards |
| Hong Kong street with Chinese/English signs | All fail | Gibberish text in all three |
| ”Gameplay of GTA 6” | GPT 2.5 | Added a logo and a wrong minimap |
| Biology worksheet labelled in student handwriting | All fail | About half wrong |
| 3×3 endemic Borneo frogs with names | All fail | |
| World topography map with labels and stats | Close, GPT models | Missing African and Southeast Asian labels |
| Room photo from the main door of a floor plan | GPT 2.5 (least wrong) | None took the right viewpoint |
| Chess mate-in-two diagram | All fail | |
| A–Z animal alphabet poster | GPT models | 2.5 adds unrequested extras |
| 11:15 on a clock, wine glass full to the brim | GPT 2.5 | Both GPT models correct |
| Where’s Waldo scene | GPT 2.5 (least wrong) | He could not find Waldo in it |
| K-pop influencer holding a text-heavy product | GPT 2.5, except a malformed hand | Text preserved; GPT 2 got scale wrong |
Variants, access and leaderboards
- Flare: faster, about 50% lower latency, slightly lower quality. Sunburst: “built for premium workflows with tighter control,” slower. Same price.
- Access (at recording): rolling out to ChatGPT, Work and Codex on all tiers including Free, with usage limits; also via the API. Up to 4K output.
- Leaderboards he shows: #1 on the text-to-image arena, “just slightly above” GPT Image 2; 2.5 Sunburst #1 on image editing, “much higher” than GPT Image 2.
Try It
- Make editable ad assets: ask 2.5 to “break this poster down into separate transparent layers, then put all the layers together into a PSD file.” Re-layout the elements in Photoshop or Photopea instead of regenerating.
- Generate layered illustrations for parallax or motion graphics: request N sequential transparent layers from foreground to background, assembled into a PSD. ^[inferred] Layered PSDs drop straight into an animation tool.
- Annotate, don’t describe: draw arrows and notes on the image itself and prompt “use GPT Image 2.5 to edit the image based on the annotations and instructions in the image.”
- Keep GPT Image 2 for literal jobs (dense grids, chart conversions, UI redesigns) where 2.5’s extra detail gets in the way, and check every number in a generated chart.
- Budget by quality setting: draft at low quality (about half a cent), finish at max (about 21¢ at 1024²).
Open Questions
- The 4K max-quality price is garbled in the caption (“roughly 40 widescreen 4K image”); confirm on OpenAI’s pricing page.
- Single generations per model per prompt; no repeated trials or blind judging.
- Which interface ran the PSD assembly (ChatGPT’s code tool or Codex) is not stated; ChatGPT presumably wrote the PSD file with code.
Related
- ChatGPT Images 2 (GPT Image 2) — Launch Coverage — OpenAI’s Images 2.5 announcement and the Flare/Sunburst split
- ChatGPT Images 2.5 — Field Test (MattVidPro) — a second creator’s test, with the same “upgrade, not a leap” verdict
- GPT Image 2.5 — Comparison Tests and Prompt Habits (AI Master) — a third test with a reusable protocol
- Nano Banana — the Google model in the comparison
- Local Image Models in ComfyUI — Qwen Image 2.1, the open model that also outputs transparency
- Awesome GPT Image 2 — prompt library for the previous model