Has the first AI imaging product emerged to receive DeepSeek dividends?

Image source: Generated by AI

After the fire of the R1 reasoning model, the National Relay integrated DeepSeek, which has a large model cloud service platform such as silicon-based mobility, a Chatbot such as Tencent Yuanbao, and even a top stream such as WeChat. However, there are few news that AI image products have access to DeepSeek R1. It was only 10 days from the release of DeepSeek R1 to Krea’s announcement of the launch of new features. This response should be the fastest among image products.

Use DeepSeek to lower the bar and

Krea took the lead in “eating crabs”

Before Krea, the function of generating and changing pictures integrated with Chatbot already existed, the most typical example being ChatGPT. However, readers who have used it should have some feelings. Although it has superimposed the steps for Chatbot to understand user intentions, whether it is generating or changing pictures, the effect is not satisfactory. (See case comparison below)

The introduction of DeepSeek-R1 has brought some changes to the experience of this type of function. Two points are in-depth reasoning and low cost.The former solves the problem of controllability, while the latter determines that new technologies can be applied to commercial products.

When users express their intentions in natural language, they will not be as clear as carefully written Prompt, but DeepSeek-R1 focuses on understanding the user’s intention.And Krea also adds buff for subsequent operations needed to get close to availability.

For example, in the above video, R1 is added to lower the threshold so that users can use natural language to generate and change pictures. Therefore, for some subsequent operations, such as changing styles and generating animations from pictures, Krea has set up convenient interactive buttons, but in terms of interaction, In fact, the product “translates” the button function into a sentence that is convenient for R1 to understand, so that the generated results are close to the user’s expectations.

The author also conducted a comparative test of generating and modifying pictures on OpenAI’s newly launched o3-mini inference model and Krea added to R1.

Photo generating instructions: Help me generate a picture of a sunset, beach, blue sea, coconut trees;

Modify picture instructions: remove all coconut tree. Note: Flux is selected by default for Krea’s generating picture model, ChatGPT selects GPT-4o ordinary model and o3-mini inference model, and the generating picture model is DALL-E3

First, the author used the same natural language instructions to ask Krea Chat and ChatGPT (two models) to generate a picture (as shown on the left), and then issued a map modification instruction to both at the same time, requesting that the coconut trees next to the beach be removed (the result is as shown on the right).

We saw that both the Krea and o3-mini models had removed the coconut trees and completed the task, but the coconut trees were still in the picture generated by GPT-4o. Compared with the two inference models, although the coconut trees were removed from the image modified by o3-mini, the image is almost equivalent to “recreating”. One of the surprising things about Krea is that only the coconut trees were removed from several trees, and most of the elements of the original image are retained, which at least can be regarded as a modification.

Behind the difference in effects is the difference between R1 and o3 inference models. In fact, there is also the difference between the living picture model Flux and DALL·E3. Since there is no way to conduct cross-verification, it is not known what factors affect the performance of different products.

This time, Krea integrated R1 into its product with another prerequisite-cost.

In the Zebra Logic test, which tests natural language reasoning, DeepSeek-R1 scores not as good as OpenAI-o1, but it scores much higher than other non-reasoning models.

Note: The o3-mimi used in the above test came online later than

DeepSeek-R1, API prices are also higher than DeepSeek

Although OpenAI o1 was the earliest inference model to be launched, due to its expensive use price, it was not realistic to improve the effect of drawing. DeepSeek-R1 completely lowered the price.

The introduction of R1 by Krea not only brings the generation and modification of maps closer to user expectations, but another significant benefit is thatLower the threshold for use, whether you no longer need to write a prompt carefully or you can take subsequent steps with a click, it is more friendly to ordinary users.

Traffic growth is stagnant, Krea is eager to break the circle

Regular readers of Beluga Whale sailing should know that we have been updating the AI Photo Navigation Series for the past six months. Krea is a product that we feel has its own unique positioning and characteristics among all photo editing products.

Data on the number of visits per capita and visit duration per capita for image editing (top)/generation (bottom) products in December. Note: Krea functions cover image generation, modification, and enhancement. It is difficult to simply classify them into the raw image/retouching category, so here we compare them with retouching and raw image products respectively| Beluga whales go to sea to map

From the perspective of the number of visits and duration per capita, Krea’s data is still good, indicating that it has gained recognition from some users. However, from the perspective of traffic, a product with good reputation and good usage data has been in a sideways position, and even declined in the second half of 2024.

Until January, Krea frequently updated and launched real-time 3D generation and rapid customized training Avatar. Two relatively broken functions, their traffic improved in January.

The reason for this is related to the fact that Krea’s relatively unique experience attracts mostly professional users (explained in detail below). As traffic declines, Krea attempts to break through the circle to find incremental users. Victor Perez, founder of Krea, also said in a podcast: “Currently, Krea’s users account for a high proportion of professionals, but as far as the team’s plans are concerned, they prefer to make AI tools that are truly for ordinary people.”

Adhering to this philosophy,Krea’s product design and iteration direction has always been committed to simplifying operations and optimizing users ‘interactive experience to attract more diverse users. The introduction of DeepSeek-driven Krea Chat is actually a continuation of this idea.

Solve the “controllability” problem by generating pictures in real time,

Krea earns its first wave of users

Krea was founded in 2022. Founder Victor Perez received millions of dollars in initial investment from entrepreneurial incubator HF0 and founded Krea. (Note: HF0 is a startup incubator located in San Francisco. They will provide entrepreneurs with a closed working space for 12 weeks and provide investment in entrepreneurial projects based on the founder’s performance during this period. The amount is more than US$1 million. There is no cap and has a 5% stake in the startup.)

Krea itself does not develop its own model,Taking the path of “open source model + optimized engineering UI”, combined with the author’s experience and the founder’s sharing, Krea’s core difference is “precise control”. In order to achieve precise control, Krea focuses on interactive optimization., and the current main real-time picture generation function can well reflect this idea.

Krea landing page

Krea began closed testing in early 2023, and the real-time map generation function was also born during the closed testing period. According to founder Victor’s sharing, the inspiration for real-time generation of pictures came from “screen projection”. He thought that when users generate pictures,The results can be displayed in real time, and then the user can follow the results and further interact with the AI to form a unique user experience.”

Statement from a Krea investor| Photo source: YC

Later, he spent the night writing the Demo and posted the test video on Twitter. Many users expressed great interest and wanted to try the program. User feedback has doubled Victor’s confidence. After several months of development, when Krea was officially launched in November 2023, the real-time generation function became the main feature. An investor in Krea said that he invested in Krea because he took a fancy to Krea’s low-latency real-time mapping.

Image source: AI Product List (December 2023)

One month after the launch of the real-time graph generation function, Krea’s website traffic increased by 191% to 3.65 million in December 2023, and the next month (2024.1), the number of visits increased by 22% to nearly 4.5 million.

Note 1: I chose a reference background here, but many creators will choose to create directly on a black background. Note 2: If you upload a reference picture or style, weight control will appear in Krea, which is a floating point value of 0-1. The closer the value is to 1, the closer the final output will be to the reference image/style.

In the time drawing interface, the canvas is divided into two parts,The left is the creative area, and the right is the finished product area, the user enters Prompt below, and the first version image is displayed on the right. In addition to Prompt, users can upload reference pictures, select a preset style (official and community-trained Lora) to control the effect of the generated picture, and adjust the “weights” of each style/image to more accurately control the output. (Krea’s Wensheng function has a similar design)

Through graphics and graffiti, you can control the shape, position, and color of important objects in the screen. Note: The higher the AI Strength, the greater and lower the weight of AI creation, and the more loyal it will be to the content drawn by the user.

Subsequently, users can “hand-draw” graffiti or insert graphics (rectangles, triangles, and circles) in the creation area on the left to control the position and shape of characters or objects in the screen, and adjust the AI Strength parameter to balance their “creation” and AI’s “play”, all changes will be displayed on the right side in real time.

According to data given by the founder of Krea, the delay of real-time display is about 40 millisecondsWhen the user drags the graphic, Krea will also continuously process the image at a rate of 4-5 frames per second (equivalent to about 250 milliseconds processing an image). The founder explained that real-time graphics generation already requires high computing power. If high concurrency is taken into account, the computing power reserve required will be even greater.With a delay of 40 milliseconds, users will not feel particularly stuck, but will also try to save computing power.

Share by a Reddit user who claimed to be a Krea employee| Image source: Reddit

In terms of engineering, Krea insisted on deploying its own models to reduce the delay caused by using third-party hosting platforms. The team also spent a lot of time developing and optimizing infrastructure such as computing power and models, allowing Krea to operate with a larger number of users., low latency can also be maintained.

After the user completes a version of the design, he can also drag the “finished product picture” on the right side to the left, and use this version of the image as the basis for further modification and optimization to achieve multiple iterations.Similar to Krea Chat, real-time graphics generation is also a variant and extension of the Wensheng graphics function. Users use graphics and graffiti to modify the image accurately according to their wishes, and can iterate multiple times to reduce the impact of the uncontrollable graphics generation model. Compared to unlimited card drawing, this mode seems more efficient.

But compared to Krea Chat, the threshold for users to control the effect of graphics through graphics and graffiti is actually not low., both the author and colleagues in the editorial department who have tried Krea have the same feeling,”I know that this real-time picture generation function is very good, but I just don’t know what it can do for me.”

From the data point of view, by February 2024 (three months after the real-time production map was launched), Krea’s traffic has exceeded 6 million, which is equivalent to that in the past year, Krea’s traffic has not exceeded this “ceiling”.

Krea creates a complete workflow for professional users,

Can it attract ordinary users?

Krea landing page

In addition to the real-time image generation and Krea Chat functions described above, Krea provides image functions such as cultural image generation, image modification, and image enhancement. In addition, there is also a function to generate video. This function was launched in May 2024 and is now connected to mainstream video generation models on the market such as Hunyuan, Runway, Kling, and Hailuo.

Wensheng pictures (top), remapping (middle), image enhancement (bottom)| Photo source: Krea

From drawing, changing to image enhancement, Krea has basically formed a relatively complete workflow.After the overall experience, the effects of these functions are very online. Continuing the overall idea, Krea has also worked hard on interactions based on precise control of excellent interactions, and the operation of these functions is relatively simple.

Users can directly box and select in the screen and enter Prompt to add objects or modify images locally

The vast majority of users on social media have praised Krea’s idea of optimizing “precise control”. This is consistent with our previous understanding. As a professional user attracted by Krea in the early days,”controllability” is just a need, and Krea also responded to this demand. Judging from the users who interact with Krea’s official account and founder Victor in X,Krea’s current user portraits are probably artists, designers, and marketers who need to make pictures frequently.

However, for ordinary users who use AI imaging products for entertainment purposes or at low frequencies, the “controllability” provided by real-time image generation may not be attractive because they do not have a figurative understanding of the final image. This is probably the reason why Krea has been difficult to “break the circle” in the past six months.

But in contrast, Krea Chat, whether it introduces DeepSeek-R1 to better understand user intentions, or uses tags to allow users to issue accurate “map change” instructions with one click,In fact, they are all responding to the needs of ordinary users for easy operation, inspiration guidance and intelligent remapping.After the launch of Krea Chat, whether this product can break through the circle and find new user growth points depends on their subsequent traffic performance.