GPT-6 Astra + Images 2.5: 一句話讓 AI 完成整個工作流！

OpenAI recently released a brand new image generation model, DALL-E 3.5. After testing it, I have no doubt that it is one of the best image generation models currently on the AI market. Even better news is that OpenAI also launched GPT-4o a week ago, so in today's video, I will demonstrate several key updates to this image model . More importantly, I will combine it with GPT-4o to show you how to use them to complete a real work task, and I will share the entire workflow with you for free. So, without further ado, let's get started. First, the new 2.5 version has now been rolled out to all users, including free users. As long as you see the update in the image section on the left side of the ChatGPT web version, it means you have access to it . On the desktop app, you can see the " Images" option on the left whether you are in chat mode or work mode. Additionally, if you want ChatGPT to generate images, you don't necessarily have to be in image mode; you can just type your request directly into the chat box, and it will generate images for you. OpenAI claims that images generated by this new model will have clearer details, more natural lighting and textures, and it handles reference images we upload much better. The biggest advantage is that you can stitch photos of yourself and your friends together without any visible traces; it looks exactly like you. You can also perform various edits on the generated images, and you can make multiple rounds of modifications to the same image, as it retains all your edit history. Although the editing performance of the previous 2.0 version was decent, if you made too many changes, it would forget earlier edits, which was quite annoying to regenerate. Besides that, OpenAI also says the model is now about 50%faster at generating images. I’m not sure if it’s truly 50%faster, but we definitely feel the generation speed has improved compared to before. Now, let's look at the biggest upgrade in this model, which is the newly added feature called "Sketch." In ChatGPT, whether you're in the chat or work interface, you can click "Sketch" under the image category, or type the "@" symbol in the chat box and select " Sketch" to draw anything on the canvas and let GPT turn your rough draft into a realistic image. Next, let me show you my real skills and my artistic talent. I can draw a cartoon sketch in under a minute. For this one, we're going to draw someone with a fierce expression, keeping it a bit exaggerated. How about that? Pretty impressive, right? That’s what you call professional. Now, you can create something like this with just a simple sketch. Of course, besides drawing, you can add a prompt in the chat box to set the style you want. And if you aren't used to drawing, you can simply describe your requirements in the chat box, which can sometimes be even more convenient. Before we move to the next feature, if you feel like you learn something new from my videos, I've actually launched the "Li Factory AI Weekly." It includes not only the week's hottest AI news but also my exclusive insights and a collection of useful AI tools I personally use, all shared for free. It’s packed with value. If you're interested, the link is in the description and the comments section. Let’s continue. Next, let’s see if giving GPT a reference image really keeps the image consistency as well as the official claim says. I’ve uploaded a photo of myself in Singapore and asked it to change my outfit to a tropical island style and turn the matcha drink in my hand into a sugarcane. I must say, this photo looks completely natural. The portrait and image proportions are kept perfectly, the island-style outfit and sugarcane look seamless, and the hand holding the sugarcane follows the laws of physics. The next example is for when you want to replace some landscaping or equipment in your backyard and want to see the effect in advance; you can use GPT to achieve that. For instance, we upload a photo of the backyard and ask it to replace the fire pit with a small fountain to see what the changes would look like. You can see this is the photo before the renovation, and this one is after. Honestly, it did a pretty good job. Aside from replacing the fire pit with a fountain, all the other elements remained unchanged. Additionally, you can upload your own photo to GPT to create a video thumbnail directly, and it really looks like me. I also had it generate photos of me wearing outfits in different colors. photos, and the consistency of these photos is very high. I must say, this model maintains portrait features much better than any previous image model. Previous models were quite different; the differences between versions were actually significant, but now those differences are very small. Beyond just generating images, the new model also offers various editing features, the first of which is the ability to adjust the aspect ratio with one click. For example, we click on the image GPT generated, select "Resize," choose the 16: 9 widescreen format, and GPT gets it done immediately. Next, let's look at "Select." This means we can draw directly on the GPT-generated image to highlight specific areas and ask it to modify them. For instance, here we ask it to increase the water flow of the fountain and add a small dog on the lawn. As you can see, GPT followed instructions and made the fountain larger, though perhaps not as large as I expected. I also realized that when using this feature, it's best to be as clear as possible with your requirements. For example, I just said to make the fountain bigger, but GPT interpreted that as making both the water flow and the fountain's base larger. Additionally, we can also add text to the images. For example, if we want to add Mid-Autumn Festival elements to this cartoon, we can draw a shape behind its head and ask it to write "Happy Mid-Autumn Festival," then see if GPT can figure out what that shape is. We send the prompt and see what result GPT generates for us. Wow, unexpectedly, it generated a moon that matches the Mid-Autumn theme. I also want everyone to notice, besides the final image generated, the changes in image details throughout the process. We started from a sketch, turned it into a cartoon, and finally added a background related to the theme. No matter how we change our requirements, the image consistency remains very high . If you've used previous versions, you know that the model used to lose elements with each new generation, making it hard to maintain consistency; however, DALL-E 3 has improved significantly in this regard. You can also try removing the background. For example, we upload a photo, select " Remove Background" here, and it generates a PNG format image. Of course , you can also use the comment tool. For instance, we can select "comment" on this cover image, then choose the area that needs modification and directly provide our revision suggestions here. For example, here I asked it to remove the avatar in the corner, change "Gpt" to "GPT," capitalize all the letters in "images," and add a GPT logo. Now, a brand-new cover image has been created according to our requirements, though for some reason, the GPT here also modified my portrait on the cover. We can also erase anything. This erase tool doesn't just wipe things away directly. Instead , it marks the area and then lets GPT delete that selected region for us. However, I've noticed that every time GPT erases something, it replaces the previous element with something else. For example, here it replaced it with paints and a pen holder; perhaps it feels that such a cover would be richer . Next, let's take a look at the template feature. Now, to help us save time, GPT has generated different templates for common daily scenarios, so you don't have to start designing from scratch—you can just pick a template and get started. Although GPT had similar features before, these templates are more professional, such as those for logos, posters, and interior design. As long as we click on a template, GPT will automatically fill in the preset prompts. Let me give you an example. I found a random house on the Australian real estate site REA, downloaded a picture of its living room , and then selected the interior design template. It will ask me a few questions first to clarify things before generating the image for us. However, at first, the result wasn't very good because what it gave me was just a single image, not the kind shown in the template. Then I took a screenshot, sent it over, and explained that this was the effect I wanted. Fortunately, it finally created it for me. While it didn't perfectly replicate the source image, there has at least been some improvement. Besides templates, there is also a "Popular" tab here, which contains even more ready-to-use templates, such as the recently popular 80s retro-style photos . Beyond helping us draw, this next question is actually more critical: can AI accept an image first and then continue with the next step of work based on it? Because often, creating and editing images is just the first step in the entire process, and what we can do next is the real key. That's why we need to use the newly released GPT-6 STR-A model. If you are a paid user, update the app to the latest version, then switch to the workspace. On the page, select it in the model category, and you can hand over control of your computer to the AI for remote operation . By the way, when using GPT-6, I personally recommend choosing medium intensity, because it strikes the best balance between performance and cost-effectiveness. Of course, you can set it higher or to ultra-high, and the results will indeed be better, but the token consumption rate is astonishing. In the demonstration case I'm about to show, we let GPT-6 automatically trigger four different apps in the workflow: first, create a brand-new image based on my photo, then make a poster in the design software Canva, and finally send that poster to a print shop to be printed. So what you see now is a very detailed workflow description I provided. In the prompt, we also specified the required tool permissions , clear stopping rules, and our expected output results. Additionally, to ensure this workflow runs smoothly, we logged into our GPT and Canva accounts along the way. So, what exactly does this workflow contain? We first authorized GPT-6 to access a folder in my Google Drive, which actually contained two of my photos. Then, in the prompt, I described the desired effect for the poster: to create a set of AI offline promotional event posters featuring me as the lead. Next, the AI imported this request into the web version of GPT and, just like we demonstrated earlier, used the latest image generation model, Image 2.5, to edit the photo and generate the first image. Then, it switched directly to Canva and handled all the poster layout work itself. I want to remind everyone again that throughout this entire process, I only gave it one set of instructions and granted access permissions; it did all the rest of the work. Finally, Extra generated two print-ready PDF versions for us. It also set the parameters in the printing service as requested, so we only needed to enter the address to pay and place the order. The whole process, including waiting and revisions, took about 40 minutes. You might say that's quite a long time, and you're right, but if we did it ourselves, it would likely take even longer. fast, but here with BT6, it actually offers us a new possibility . Before, you might have used AI to create an image, then put it into software to edit it, and finally printed it out. But now, AI can not only do these things, the key is that it operates autonomously, and while it works, we can still edit the content, undo any step, or make other adjustments at any time without breaking the workflow. I think future AI will become more and more like a colleague working right beside us, and achieving all of this might just require a single command from you. Of course, the poster creation we demonstrated today is just one of many possibilities. You can actually use this methodology to help with research, prepare weekly briefings, or write scripts. I even think it will soon be able to help us generate audio, edit content, render videos, and handle final quality control, while all we provide is the initial creative idea and the final judgment and fine-tuning of the output. Another example is that you can have AI handle your weekly operations. For instance, you can ask it to pull information from one app, financial data from another, and next steps from your meeting notes, then finally have the AI consolidate everything into a weekly operations report, helping you list task priorities, owners, bottlenecks, and potential solutions. This way, you only need it to generate one report per week to stay updated on how things are progressing. To be honest, these tasks used to depend on project managers, but in the future, you might be able to use AI to do them instead. The key is that it can also maintain the consistency and stability of the task from start to finish. So that is the new GPT-2.5 image generation model. I really like it because it can already produce some great images, gives us more control, and handles reference images and editing better than any previous model. Of course, the current version 2.25 is still not perfect and has various issues. For example, while it is quite good at maintaining consistency for simple images, there is still room for improvement when it encounters truly complex ones. Additionally, when it removes certain elements from an image, it might add some other elements on its own, and in terms of following our instructions, it still requires us to be very detailed. Also, when it comes to generating photos of celebrities or copyrighted intellectual property, GPT can't do that for the time being. But when we combine the image generation feature with GPT-6, the possibilities multiply, because we no longer need to download images and move to other software; the images generated right here in GPT can serve as the raw material for our next task. We can also use GPT-6 to call multiple applications simultaneously, helping us get the actual results we want, rather than just the first step of a long process. Although it’s not fully automatic right now—we still have to log in and authorize it—once you manage to set up these workflows, you can have AI working for you in the long run. This new image generation model is really impressive; while it's not perfect, I will definitely be using it often, and it will be my primary model for generating images from now on. OK, that's all for this video. If you found it helpful, please like and subscribe to the channel. See you in the next one , bye!