preamble
The night before Halloween October 30, an operating friend told me a point Halloween avatar generator, and then probably gave me an analysis of the overall idea, so I used the button Coze platform () to build an AI intelligent body The whole process took an hour to get it done! I deployed a key to my AI small program, the next day randomly posted a small program to visit the page data directly increased by 1000%, then I came to disassemble the whole process.
Those who don't know about the Buckle Coze platform can this piece ofTeach you to build AI apps in 5 minutes (no coding required)》
disassemble
On the evening of October 30th, I received a message from a friend.
Then I experienced the overall need to collect the user's avatar, gender, drawing style, Halloween elements, and then generate a Halloween avatar.
When I experienced it, I felt the effect was quite good, and the friend also analyzed the whole idea of the implementation and the cue words for AI painting.
I did some interaction optimization based on this idea, for example, her case is to collect the user's avatar, and then let the user to choose the gender and painting style and elements, this form belongs to the way of form collection, but the AI intelligences are in the form of dialogue, if to collect so much information in the form of dialogue, it's very cumbersome for the user, so I changed to only need the user to send a photo, the gender is extracted from avatar inside, and then the painting style and elements are randomized.
There are two advantages to this:
-
Less user resistance and lower barriers to use
-
Generate more variations of results to add interest
Once the outer logic is confirmed you start building the entire workflow, which is very simple with only 5 nodes.
first message node
Output randomized styles and elements + loading animations to give users something to look forward to making it easier for them to wait.
The second "Image Understanding" plugin node
To extract the features of user uploaded photos in order to subsequently generate image cues, the visual understanding of the "Step Star" large model is used, which recognizes images with high speed and accuracy.
Large models of visual understanding: /docs/guide/image_chat
PromptText: The picture is a photograph of a person and you are asked to describe TA's skin color, hair, eyes, nose, mouth, face shape, facial features, national or ethnic characteristics, height, build, age, expression, makeup, glasses, earrings, actions, clothing features, and accessories.
Output Content Reference Case Format: Gender: Female, Face: Melon Face, Body Type: Slightly Chubby, Age: Around 40, Appearance: Asian Female, Skin Tone: Light Appearance: Long black hair, wearing a black newsboy cap Appearance: Very calm, with the corners of the mouth slightly raised Makeup: Light makeup was applied Actions: Reaching out to gently caress the horse's nose Clothing: Wearing a yellow long-sleeved shirt, dark pants, and a slung black Wearing a yellow long-sleeved shirt, dark pants, and a black Champion bag.
The third "Text Processing" node
The main thing is to assemble the cue words for AI painting to generate pictures, combined with random styles + elements + character traits after picture comprehension.
The fourth "Image Generation" plugin node
The last assembled cue words to the AI painting for image generation, here using the "Step Star" of the Vincennes big model, generating images very quickly, and the quality is quite good.
Vincennes model: /docs/guide/image_generate
The fifth message node
Output the results of the generated image display, and finally prompt the user to generate different each time to guide the user to generate again.
After this AI smart body is built, our team built a set of small programs to access the Buckle API, and we only need to publish the smart body into our AI small program through BotId in the background, without the need to develop and send out the version.
Buckle API: /open
summarize
This hot pursuit reminds me of a book I recently readLook at the case of the collection of formulas mentioned in Addicted:
Trigger (holiday hotspot triggers) + action (uploading an avatar generates) + variable reward (randomized combo results) + commitment (wanting to see another combo) = addictive!
There is also the point that, as stated in theWhat can I write about?A point was made in the article:
I think when you want to make a product, a lot of the functionality doesn't have to be solved by writing code, it can be solved by using plug-ins and workflows from the Intelligent Body Building Platform.
With the code development function may all need a day to complete, and in the platform intelligent body building an hour to get it done, and then expose the API out can be very quickly complete this product.
I think AI Intelligent Body + AI applet = MVP (Minimum Viable Product) king bomb combination, it can effectively reduce the cost of trial and error.Saving money when you're not making money is the way to go.。