Google's New AI: Gemini Canvas & Consistent Image Models

TL;DR

Google's new AI tools can now create consistent characters across multiple images and let you build websites in real-time with Canvas. #VentureStep #GoogleAI #Gemini

INTRODUCTION

Creating compelling and consistent visual assets has long been a significant hurdle in marketing and design. How do you tell a continuous story when your AI image generator produces a different character every time? Similarly, how can you rapidly prototype a website or document without constantly switching between a chatbot and an editor? These challenges slow down creative workflows and create frustrating inefficiencies.

In this episode of Venture Step, host Dalton Anderson dives into Google's latest AI releases that directly address these problems. Setting aside recent news from other players, Dalton focuses exclusively on the new capabilities within Google's ecosystem, showcasing tools that are pushing the boundaries of what's possible with generative AI for practical business applications.

Dalton explores two groundbreaking features: an experimental image generation model, Flash 2.0, which can maintain a consistent subject across multiple prompts, and Canvas, a new function within Gemini that allows for the real-time creation and editing of websites and documents. He walks through several powerful examples, demonstrating how these tools can be used for everything from product modeling to instant document creation, offering a glimpse into a more integrated and efficient creative future.

KEY TAKEAWAYS

• Consistent Subject Generation is Here: Google's new AI model can maintain a consistent character or object across multiple images, a game-changing feature for storytelling, marketing campaigns, and product design.
• Prototype Websites and Docs with Canvas: Gemini's Canvas feature enables users to create and edit websites, app mockups, and rich-text documents in real-time directly within the chat interface, streamlining the development process.
• Powerful, Context-Aware Image Editing: The AI can understand complex prompts to modify existing images, such as correcting a person's posture or generating realistic hands that weren't in the original photo.
• Practical Applications for Business: These tools offer immediate value for creating marketing materials, modeling products with consistent subjects, and rapidly generating formatted business documents.
• Experimental Tools Come with Bugs: While powerful, new features like Canvas are still in their early stages and may have bugs; Dalton specifically notes that the versioning or "undo" function can break the prompt sequence.

FULL CONVERSATION

Dalton: Welcome to Venture Step podcast, where we discuss entrepreneurship, industry trends, and the occasional book review. ¹Unfortunately, I had some audio issues last week, so I will be re-recording this episode. ²Hopefully, when you go to look at the most recent episode and the previous episode, they both have audio. ³If you were trying to listen to the episode last week, it was because I was having audio issues, and unfortunately, during the week, I was just slammed and could not re-record. ⁴So here I am recording the same episode that I did last week. ⁵I hope that you'll enjoy it this time when it has sound. ⁶

Apologies and Episode Overview

Dalton: So what are we discussing in this episode? ⁷We're pretending like there wasn't a recent release of ChatGPT's image model. ⁸In this world, I don't know this information. ⁹And it's really just talking about Google and some of the Google releases that they've had. ¹⁰Of those releases, they had some really cool releases related to image generation and native image generation. ¹¹

...the model understands the context of the image and can have a consistent character use throughout your image creation. ¹²

Dalton: Not only can you create images, but the model understands the context of the image and can have a consistent character use throughout your image creation. ¹³So if you provided a character, you can tell a story with it. ¹⁴You can make marketing materials. ¹⁵You could do product designs. ¹⁶You can make models. ¹⁷And so that is something that you couldn't do before, especially natively in an app. ¹⁸Keep in mind that you can only do this in what they call Google AI studio. ¹⁹

Dalton: They also released another thing called Canvas, and Canvas is supposed to be used for generation of documents and websites and previews of apps that you create. ²⁰This isn't a particularly new feature. ²¹Anthropic has had their version of Canvas, which they call artifacts, for quite a while. ²²So don't think it's brand new, but I do think the take of rich text editing within a document, within your chat, is pretty interesting. ²³

Dalton: That's what we're going to be going over today. ²⁴And I hope once again that this episode has sound. ²⁵I tested it a couple of times, so we'll see. ²⁶Fingers crossed. ²⁷

Native Image Generation with Consistent Characters

Dalton: All right, so the first thing that we'll be discussing is the native image generation. ²⁸I have some examples that I'm going to put on the screen. ²⁹If you're listening via audio only, that's fine. ³⁰I'll narrate a little bit of what's going on and you'll be able to understand. ³¹

Example 1: Correcting Posture Instantly

Dalton: Okay, so in this first example, it has a model that is slouched over, not looking very confident, and she's got really bad posture. ³²This could be from a disability or this could just be because she's not having a good day. ³³Say that you wanted to change your image and make yourself have good posture. ³⁴There's a side-by-side where one person is doing all the work that would be required to do this in Photoshop. ³⁵And then there's another version to the right of the screen that just says, "Make the girl stand in correct standing posture." ³⁶And then it just comes back six seconds later and she's standing up straight with good posture. ³⁷So that's one example. ³⁸I thought that was interesting. ³⁹

Example 2: Creating Multiple Views of a Product

Dalton: This one I really like, and this emphasizes the consistent subject model that's created. ⁴⁰The first prompt was to create a transparent futuristic vehicle. ⁴¹It has these big off-road tires and it's got this futuristic look. ⁴²It's transparent so you can see all the parts in it. ⁴³It looks really cool. ⁴⁴The next thing that was prompted was, "Okay, now create different perspectives of this subject." ⁴⁵And so it does a front view, the three-quarter view, and then does the side view. ⁴⁶It does a really good job. ⁴⁷This one's my favorite. ⁴⁸It looks sick. ⁴⁹

Example 3: Generating Hands from a Selfie

Dalton: This one is of a woman and there's an original selfie and she looks like she's in some kind of library, I think. ⁵⁰It's a selfie of her and her arms are out, so you can't even see her hands. ⁵¹The only thing that is sent in the selfie is her smiling, her hair, and the top that she's wearing. ⁵²So then it's prompted, "Make her create a heart shape with her hands." ⁵³And then she does that with the hands that didn't exist in the selfie. ⁵⁴And they even gave it French tips. ⁵⁵The AI model gave the woman in the image French tips. ⁵⁶It successfully makes the heart shape, generates hands that don't exist, and it matches her skin tone pretty well. ⁵⁷

Dalton: Then the next prompt was, "Make her give a thumbs up," and the thumbs up works pretty well. ⁵⁸The hands are once again generated because there are no hands in the original selfie. ⁵⁹It's just using the tone of the body to figure out what the inside of your hand should look like. ⁶⁰It does a great job. ⁶¹

Example 4: AI-Powered Storytelling

Dalton: This next example emphasizes telling a story. ⁶²And I see this as a marketing plan. ⁶³

All of marketing and sales is really telling a story. ⁶⁴

Dalton: A bigger piece is community building, having in-person events, and also building in public, and also creating this compelling marketing campaign where people want to know what happens next. ⁶⁵

You can create those things now with some prompts. ⁶⁶

Dalton: The original prompt is saying something like, "I want a scene of a lonely man on Pluto, imagining a happy life." ⁶⁷So it makes the first frame of this lonely man in the middle of nowhere. ⁶⁸And then it creates another frame of the same man, and now he's holding hands with his partner. ^69696969He's eating dinner and it's the same person. ⁷⁰It's five shots. ⁷¹And then it flips back to where he is, actually on Pluto in this made-up story. ⁷²It keeps going back and forth between the different memories. ⁷³Compelling stuff. ⁷⁴

Example 5: Modeling Products with AI

Dalton: This is also one of the key things that you can do if you can have a sustained subject in your image generation: you can use it to model products or you can have your products being used by models. ⁷⁵

If you have the same subject modeling the item in different ways or using it different ways, then it's more compelling. ⁷⁶

Dalton: In this example, it says, "Create image, make the girl in the photo wear the jewelry in the second photo." ⁷⁷It's this expensive-looking piece of jewelry, a gold emerald necklace with pearls and gold emerald earrings. ⁷⁸And then the image generates an image of the model wearing the jewelry that was required, which is great. ⁷⁹

Dalton: Here's another really good one. ⁸⁰The original image is a woman wearing a top and some jeans, and she's smiling. ⁸¹And from that, they created four reference images for the model. ⁸²One is with her smiling, looking to the side; one is with her being playful; another of her smiling and drinking a cup of coffee; and another one from a side angle where she is smirking. ⁸³You could take a model, hopefully with their permission, and you can provide an image of an object that you want them to be wearing. ^84848484Then from there, you can create these consistent subject variations of your original idea. ⁸⁵

Introducing Gemini Canvas for Prototyping

Dalton: Okay, so the next thing that we're going to touch on is Canvas. ⁸⁶Canvas is within Gemini, at gemini.com. ⁸⁷And Canvas, as I mentioned, isn't a new idea. ⁸⁸Other companies like Anthropic have had artifacts. ⁸⁹I do think it's a great addition to Gemini and it allows for quick prototyping of websites or apps or creating documents and then editing those documents in real-time with your adjacent AI buddy. ⁹⁰

Live Demo: Building a Website with Canvas

Dalton: So I created a very simple website using a prompt that I created earlier. ⁹¹I'm going to be creating a maritime insurance coverage website. ⁹²I'm going to select Canvas. ⁹³Canvas opens up and takes up about two-thirds of your screen and creates this website. ⁹⁴So in this prompt, I said, "Can you help improve the website, add whatever you want? I'm presenting this to my boss, asking for a demo on Monday." ⁹⁵They replied back that they want to do content expansion, enhanced styling, and responsive design. ⁹⁶

Dalton: All right, let's see. ⁹⁷I can see that the website looks visibly better. ⁹⁸They added a solid header styling. ⁹⁹When I click on the headers at the top, like Cargo Insurance, Hull Insurance, it brings me exactly to that area on the website, which I think is great. ¹⁰⁰So I'm happy with this. ¹⁰¹I'm sure my imaginary boss would love it. ¹⁰²

Live Demo: Creating a Document in Canvas

Dalton: So this one, I want to create a document. ¹⁰³It's the same gist where you click Canvas and Canvas creates a website, an app, or documents for you. ¹⁰⁴I already created a wonderful outline for AI to process and turn into a document that I could share. ¹⁰⁵Okay, so it breaks down the different types of coverage in maritime insurance. ¹⁰⁶We have transport cargo, hull insurance, offshore energy, marine liability. ¹⁰⁷But one of the things that it doesn't show is what are things that typically aren't covered. ¹⁰⁸So let's add that. ¹⁰⁹

Dalton: Okay, so it said it added a common exclusion section. ¹¹⁰Now I'm asking it to add the exclusions to each section. ¹¹¹Okay, so it updated the document. ¹¹²This is exactly what I'm looking for. ¹¹³Each section of the coverages now has common exclusions. ¹¹⁴For transport cargo insurance, exclusions often include inherent vice, spoilage of perishable goods, improper packing, and delay. ¹¹⁵If we go to hull insurance, it's wear and tear, gradual deterioration, and damage due to lack of maintenance. ¹¹⁶

A Word of Caution on Canvas Bugs

Dalton: You can export in Docs if you wanted to change the formatting. ¹¹⁷

This canvas feature allows you to edit the documents and allow you to format it in the manner that you want... in real time, which is something I have not seen before. ¹¹⁸

Dalton: And I think it works pretty well. ¹¹⁹It has the same issue with the versioning though. ¹²⁰The versioning seems to be broken on Canvas and is something that they need to work through because it doesn't seem to work. ¹²¹I had issues when I originally recorded this episode where when you try to undo and then re-prompt, it would break the prompt. ^122122122122I couldn't prompt anymore. ¹²³So I think that's just a bug. ¹²⁴I would only use the previous version button when you're completely finished. ¹²⁵

Final Thoughts

Dalton: So that was Google's recent release. ¹²⁶Once again, sorry about the audio issues that I was having last week. ¹²⁷Hopefully, that doesn't happen again anytime soon because it is definitely a pain to re-record an episode. ¹²⁸I hope that you enjoyed this episode and appreciate you listening in every week. ¹²⁹Wherever you are in this world, have a great day. ¹³⁰Good morning, good evening, or good afternoon. ¹³¹Thanks for listening. ¹³²Hope you listen in next week. ¹³³Next week we'll be discussing OpenAI's recent release of their model and the things that you can do with that one. ¹³⁴Have a great day. ¹³⁵

RESOURCES MENTIONED

• Google AI Studio
• Gemini (gemini.com)
• Canvas (Google Gemini feature)
• Flash 2.0 Experimental
• Anthropic
• Artifacts (Anthropic feature)

INDEX OF CONCEPTS

Dalton Anderson, Venture Step, Google, Google AI Studio, Canvas, Gemini, Flash 2.0 Experimental, Anthropic, Artifacts, Photoshop, Pluto, maritime insurance, cargo insurance, whole insurance, offshore energy insurance, marine liability insurance, OpenAI