Four tests instead of demo hype: text, supposed web search, manga with a consistent character and the difficult deepfake question.
The video responds to a major product claim and turns it into four concrete tests. Instead of showing only pretty outputs, Lukas checks text rendering, research context, character consistency and safety boundaries.
The most interesting part is not only that text in images improved. It is also that a result can look convincing even when the underlying research did not happen or cannot be verified clearly.
For creators and businesses, that matters: AI image generation is becoming more useful, but the need for review, context and responsibility rises with it.
| Aspect | In the video | Why it matters |
|---|---|---|
| Test 1 | Text in image | Book title, umlauts and special characters as a realistic check |
| Test 2 | Infographic | Whether real web search clearly improves factual output |
| Test 3 | Manga | Several panels with a recognizable character and story arc |
| Test 4 | Fake-news visual | How convincing synthetic news images can look |
| Strength | Visual quality | The outputs are more useful than older models |
| Risk | False confidence | A good image does not prove the facts are right |
The video treats OpenAI's image model as a practical test: generating images, rendering text, building visual stories and probing limits.
The test shows major progress. Review is still necessary, especially for names, special characters, numbers and layout.
Because an image can look very convincing even if the factual basis is unclear. Visual plausibility is not source verification.
Better image models can also create more believable fake scenes. That makes responsibility more important for content and communication.
There is no manual English subtitle track for this video. This page uses an editorial summary instead of publishing automatic captions as a transcript.
The opening takes a large product claim seriously and tests it. Instead of repeating launch hype, it turns the claim into concrete checks.
Lukas explains what “tested” means here. The point is traceable tasks, not cherry-picked beautiful examples.
The first test checks text rendering with a real book title. Umlauts, special characters and layout reveal whether the model became practically better.
The second test asks whether the model really researched. The infographic looks convincing, which makes the factual basis even more important.
The manga test checks character consistency and story across several panels. That matters for creators because image models move toward visual storytelling.
The deepfake section is the necessary downside. As outputs become more realistic, boundaries, review and responsibility matter more.
The verdict separates progress from trust. The model is clearly stronger, but good-looking results do not prove factual accuracy.
The conclusion opens the next question: how to use image models productively without losing review and context.
I work 1:1 with freelancers, consultants, coaches and small teams on practical AI workflows, automation and custom tools. The first call is free and takes 15 minutes.
Book a free strategy call