NemoClaw Knowledge Wiki

New image generator in chatgpt

3 min read

https://www.youtube.com/watch?v=jFJAIOAzFqQ Here is a summary and breakdown of the video review regarding OpenAI’s new “ChatGPT Images” model. Greg Isenberg channel

Video Review: Taking ChatGPT’s New Image Model for a Spin

Presenter: Greg Isenberg Objective: To test OpenAI’s newly launched image model, compare it to his current “daily driver” (referred to as Google’s Nano Banana Pro), and see if the output is high-quality enough for professional use.

🧪 The Experiments

The presenter walks through four distinct tests using the preset styles available in the ChatGPT image interface.

1. The “Plushie” Test

  • Subject: A photo of Sam Altman.
  • Process: Selected the “Plushie” style from the carousel and uploaded the photo.
  • Observation: The model automatically generates an “optimized prompt” describing the image in detail (similar to how the app Glif works).
  • Result: The output was described as “super cute” with impressive attention to detail—specifically noting that the plushie retained the specific wave and style of Sam Altman’s hair.
  • Business Insight: He notes a massive opportunity for creating CPG (Consumer Packaged Goods) brands and physical toys using these models.

2. The “Sketch” Test (Graphite Pencil)

  • Subject: A photo of the presenter drinking a martini.
  • Initial Result: A detailed 3D graphite sketch.
  • Critique: While the drawing was amazing, it included a hand drawing the sketch which looked “AI-y” and unnatural.
  • Refinement: He asked ChatGPT to remove the hand and the notebook and just show the drawing on a piece of paper.
  • Result: The model successfully followed the negative constraints. The presenter noted that the new model handles feedback and iteration much better than previous versions.

3. The Diagram Transformation

  • Subject: A digital chart/diagram about “putting ideas on the internet” that he previously posted on X (Twitter).
  • Goal: Convert a clean digital graph into a “casual, natural hand-drawn” style.
  • Instruction: “Make this hand-drawn… a little more casual… remove weird pencil sharpening stuff.”
  • Result: “Wow.” The model produced a clean, legible, hand-drawn version of the chart.
  • Takeaway: This is highly useful for content creators. Hand-drawn diagrams often perform better on social media than polished digital graphics.

4. The “Bobblehead” Test

  • Subject: A selfie of the presenter.
  • Instruction: Create a bobblehead, but not in a baseball uniform. Instead, style it as a “Tech YouTuber.”
  • Result: The model nailed the outfit, capturing his specific long-sleeve sweater and adding a camera accessory.
  • Business Insight: Another potential use case for custom manufacturing and Shopify stores.

📝 Analysis of OpenAI’s Announcement

The presenter reviews the official blog post to highlight specific technical improvements:

  • Editing Capabilities: The model excels at adding, subtracting, and blending elements (proven by the sketch test where he removed the hand).
  • Creative Transformations: It can change texture and layout while preserving the original subject’s details.
  • Instruction Following: The model is more reliable at following complex grid instructions (e.g., “Draw a 6x6 grid with specific items”) compared to previous versions.
  • Text Rendering: Significant improvements in spelling and text generation within images.

🏆 Final Verdict

  • Expectations: The model “beat his expectations.”
  • Comparison: He considers it as good as, if not better than, his previous favorite tool (Nano Banana Pro).
  • Conclusion: The outputs are high-quality enough to be used for real business applications, social media content, and product design. The ability to iterate and refine images via chat is a standout feature.

Graph View