AI Card Bao HERMES.AICARDBAO.COM

Industry News

ChatGPT can now generate GIFs directly. We can use it to create emojis first, and then deal with product animations at a later time.

“Create a GIF showing a dancing cat.” This is an example that OpenAI provided for Images 2.6. In short, it allows you to generate a looped animated video, and you can also use ChatGPT to adjust the playback speed, edit individual frames, or share the video directly.

It’s really fun to play with it. If I really want to use it to promote my brand account, there are a few things I need to take into consideration first.

Before you start moving, decide on the front, side, and back first. Schematic diagram.

About two seconds – shorter than expected.

By default, the API uses 9 frames, with each frame lasting 200 milliseconds; therefore, the entire loop takes only 1.8 seconds to complete.

The official documentation doesn’t provide a consistent answer regarding the upper limit of the number of frames. The generation guide states that the range is 2 to 64 frames, while the API reference mentions 2 to 16 frames. Some developers have tested and found that attempts to use 17 frames result in immediate rejection; therefore, it is safer to assume a maximum of 16 frames at present. Although setting each frame’s duration to 2000 milliseconds would certainly allow for a longer playback time, such a setting would make the presentation more akin to a slide show rather than a fluid video sequence.

It is not possible to set both the frame rate and the image dimensions to their maximum values simultaneously. On the web version, some users request a resolution of 512×512, but what they actually get is 358×358; however, using the same parameter settings through the API, the resulting resolution is indeed 512×512.

Additionally, please note that the editing interface is not capable of generating GIF files; only the generation interface can do so. If you wish to use existing images to create animations within the API, you should thoroughly understand the relevant process in advance.

The parts that are most likely to reveal flaws are the seams.

In a set of test results released on the same day, it was found that generating 10 frames of animated GIFs using this API took 17.6 seconds and cost approximately 0.056 US dollars. Three issues were identified: the step length of the animations was inconsistent; there was an unnecessary white mark on the badges; and the poses in the first and last frame did not match, causing noticeable jitters when the animation played in a loop.

Such animations, when used in chat emojis, are completely harmless and just meant to bring a bit of fun. However, when applied to product animations on a brand’s homepage, they can be quite distracting.

So, I will first use it to create emoji, mascot animations, short promotional videos for social media platforms, and rough storyboards. For product animations and walkthrough demonstrations, I won’t use it yet. For tasks that require more than three seconds of duration, it’s better to opt for video models instead of trying to use GIFs.

The little fox that moves only one hand. This image is composed of nine static frames stitched together to illustrate how the movement should be depicted; it is not a native GIF produced by the model itself.

When writing prompt words, one should focus not only on the dynamic aspects but also on those that remain static.

These various ways of writing were frequently mentioned in the sharing posts of early users; you can definitely give them a try.

“Move only one part of your body,” the little fox said, waving its right hand while keeping the rest of its body motionless. This approach was far more stable than “the little fox greeting someone happily.” The latter option allowed the model too much room for variation—its tail, expression, and posture could all change at the same time.

Clearly state in the requirements that the loop should be implemented such that “the final frame returns to the same pose as the first frame.” If this is not specified, it is very likely that the animation will only continue until the end of the movement.

When starting from a static image, it is essential to adhere to the constraints of “maintaining the original character design” and “not changing the background.” These two principles should always be taken into consideration.

When making changes, retain the existing actions. For example, “Keep the same movements; just change the expression to a smiling face.” If only one specific frame has an issue, it is more efficient to identify that frame and make the necessary adjustments in ChatGPT rather than re-recording the entire sequence.

Adjust the speed at the very end. In ChatGPT, it can be set between 0.5 and 2 times the original speed; there is no need to regenerate the content entirely just to change the rhythm.

The character hasn't been finalized yet, so there's no need to rush into making it active.

When the animated video is played, every issue related to the character design becomes apparent. On a static image, it might not matter if the color of the scarf is slightly off, but when shown in nine consecutive frames, the changing colors will be easily noticed by everyone.

This is when the function of “saving reference images” comes in handy. However, there is currently not enough third-party data to determine how reliable this feature actually is. There is a simple way to test it: have the same character appear in five different scenes, with edits made in between. If the appearance of the character in the fifth scene is still identical to that in the first scene, then the function can be considered effective.

A more efficient approach is to separate these two steps. First, define the character and its still images on PixPix’s GPT Image 2.6 page: upload a sketch or reference image, and then use the Sunburst tool to make adjustments in multiple rounds, modifying only one part at a time. Once you are satisfied with all three views and the main expression, export the final result. The advantage of this method lies in its ability to support multiple rounds of editing and ensure consistency in the appearance of both the character and the product, making it ideal for refining basic materials that will be used repeatedly.

It should be noted that the PixPix page currently focuses primarily on the generation and editing of static images; there is no mention of GIF functionality on this page. Therefore, for creating animated videos, you can return to ChatGPT, upload the finalized character images, and follow the instructions provided above to create animations. If you need longer video segments, you may want to check out PixPix’s AI video section.

Once the characters are properly defined, adding ten more expressions later on simply involves changing the corresponding prompts.

Given its current level of development, the value of the GIF function lies not in the quality of the resulting videos, but in the fact that it reduces the cost of “testing something out first” to just a few seconds and a few cents. In the past, one would have to wait until designers scheduled their time to find out whether a mascot’s movements looked smooth or whether the rhythm of the animation was correct; now, a rough version can be seen even before a meeting is held. For testing different directions or ideas, this is more than sufficient.

Related