Article title: When Kling Image 4.0 Comes Up with Design Proposals on Its Own, the Bar for Image Generation Has Shifted Again!
Body:

Over the past few years, AI image generation models have continually reshaped people's expectations when it comes to realism, resolution, and lighting and shadow rendering. But when these images are actually used in commercial scenarios like posters and advertisements, they still face problems such as whether text can be rendered accurately and whether local edits can preserve the original design.
These seemingly trivial requirements remain difficult problems for image models to this day. (Leiphone (WeChat official account: Leiphone))
In March of this year, Microsoft Research released a commercial visual generation benchmark called BizGenEval, testing mainstream models with 400 prompts and 8,000 check questions. The results found that models can already generate fairly aesthetically pleasing posters and web pages, but still have obvious shortcomings in precise layout, attribute correspondence between multiple objects, and other aspects.
It was precisely at this stage that Kuaishou's Kling, which gained fame through video generation, today launched Kling Image 4.0.

Generated by Kling Image 4.0
According to the official introduction, the new model raises native resolution to 8K, supports up to 2,000 characters per image, and can generate up to 10 coherent images at once. More noteworthy is that it adopts a unified autoregressive architecture for understanding and generation, attempting to let the model handle image understanding and generation within the same context.
This time, Kling clearly doesn't just want to make images prettier; it also wants the model to handle more complex creative requirements.
Let's start by testing Kling Image 4.0's capabilities with the most intuitive portrait photography—from taking a good photo, to designing a poster, to organizing an entire set of visual content. How much work can Kling Image 4.0 actually do for creators?
From Portraits to Creative Posters
How Strong Is Kling Image 4.0's Image Generation Capability?
In this experience with Kling Image 4.0, what surprised us most was the realistic texture of the images it generates. We tried several portrait photos, and whether it was the subject's skin texture and details or the outdoor lighting and shadow projection, the results were already quite close to real photography.


This kind of realism has already begun to affect the photography industry itself. In April of this year, Vogue magazine interviewed some commercial photographers and found that some clients have started using AI-generated images as shooting references, asking photographers to recreate those scenes in reality.
However, portraits mainly test the model's reproduction of people and real scenes. If you want to create a poster that is actually used for dissemination, things become considerably more complicated, especially once Chinese text is added.
Recently, a conversation between guests Xu Fei and Li Na on a variety show on Mango TV accidentally made "hua di wei lao" (drawing a circle on the ground to confine oneself) a new internet buzzword. We decided to let Kling Image 4.0 join the fun too.

Compared with ordinary text posters, "hua di wei lao" is actually a very suitable subject for visual design. It has both clear textual content and imagery that can be expressed through visuals. You can see that the final generated work has quite the feel of a Mango TV variety show promotional poster: a semicircular sea-surface composition, a cool and detached color palette, and text layout settings and artistic font design that are almost at the level of a publishable poster.
There is one change worth mentioning separately: Kling Image 4.0's image editing can directly select existing text in an image for modification, and can also add new text to an existing image. The 3.0 stage already had local re-editing capabilities, and 4.0 makes the control granularity finer. We tried modifying the poster's text content and adding a line of subcopy, and only the text we requested was changed in the original poster; other layers of the image were not redrawn.

Font control is also where 4.0 and 3.0 diverge this time. At the 3.0 stage, creative fonts mainly relied on prompts to describe the style—it could be done, but controllability was limited; 4.0 makes this more tool-like: it has 70+ built-in preset fonts that can be selected directly, and also supports uploading font reference images so the model can generate according to the glyph shapes and stroke styles in the reference image.

Text generation has always been a relatively difficult part for AI image generation models. Artistic font design goes a step further than simply writing Chinese characters correctly.
For example, in the summer camping poster we made below, the artistic font blends very naturally with the scene. From the scene and color scheme to the title design, it has a fairly complete finished-product feel; with a little checking and adjustment, it could be used as an event promotional image. The font itself can become part of the画面—it can interact with clouds, waves, and other visual elements. For the model, this requires handling the relationship among text content, spatial position, and visual style at the same time.

In addition to creative posters with relatively little text like this, Kling Image 4.0 also claims to support up to 2,000 characters per image. Short titles test fonts and creativity more, while long text tests text accuracy, information hierarchy, and layout capability more. We will also continue to observe its performance in handling more complex graphic and textual content in the product detail pages later.
A New Product Launch Event for a Five-Yuan Roasted Sweet Potato
AI Even Came Up with the Copy
If artistic-text posters mainly test the model's aesthetics and creativity, then product detail pages have to handle much more complicated matters. Product photography, how to present content in the detail page, and the ordering and arrangement of images and text—these are all real commercial design scenarios.
This is also one of the capabilities Kling Image 4.0 strengthens this time. The new model supports more complex mixed image-text layouts, can hold up to 2,000 characters in a single image, and can create long e-commerce detail images in one click.
We chose an utterly ordinary product: a roasted sweet potato. We wanted Kling Image 4.0, based on this photo, to design e-commerce promotional material for a roasted sweet potato, including both product photography that attracts consumers and selling-point introductions, consumption scenarios, and necessary product descriptions.
Prompt: Please, based on the real roasted sweet potato photo I uploaded, design a complete long e-commerce detail page image for this ordinary roasted sweet potato priced at 5 yuan, with the texture of a new product launch event from a top international tech brand (such as Apple).

The result the model delivered really does look somewhat like a mature food brand.
It designed a brand identity called "ROOTS 根薯系列" for the product, chose green and off-white as the main visual elements, and used "红褐焦脆,金黄流心" (reddish-brown and crispy, golden and molten inside) as the main title. The originally ordinary roasted sweet potato was placed into a carefully arranged product photography scene, with the warm yellow flesh and dark brown skin becoming the visual center of the image.

More interestingly, the model didn't pile all the content into the main product image, but continued to extend into different types of pages: one showing consumption scenarios with sliced plating, one telling the product's origin under the theme "the terroir of sun-facing hills," and another organizing different consumption states and storage and reheating methods into an illustrated guide.

Looking back carefully, the prompt was only a simple sentence, and we didn't even provide complete product information. Yet the model proactively filled in information such as the brand name and advertising slogan it was asked to design, and organized them into a visually consistent set of works.
This was also the most surprising part of this test.
In the past, using AI image generation often required first figuring out what the image should look like, then describing it to the model in as much detail as possible—such as product position, background color, text layout, and other information. If you wanted a whole set of promotional material, you also had to design prompts one by one and repeatedly adjust the style. But this time, all we provided was one photo and a relatively vague goal. As for the more detailed visual content, Kling Image 4.0 was already able to complete a considerable part of the ideation on its own.

Behind this there is another change worth noting: AI is beginning to become familiar with the expressive methods of commercial design.
For small merchants without a professional design team, or operations staff who need to quickly produce marketing materials, this capability has fairly direct value. In the past, they might have had to separately prepare product photos, copy, and layouts; now they can first generate a visual proposal with a single natural-language sentence, then adjust it according to actual needs.
Of course, the generated results still have room for improvement. For example, the clarity of some small explanatory text is not as good as that of large headlines, and the place of origin and product information that the model adds on its own may not always be accurate. But judging from the overall effect, Kling Image 4.0 has already demonstrated a fairly complete ability to organize commercial graphics and text. Compared with simply generating a beautiful product image, it can further handle brand style, advertising copy, scene presentation, and page layout, quickly turning a simple idea into a concrete, visible design proposal.
From one image to a set of images
AI begins to learn how to tell stories
In recent years, image models have begun to explore series-based image generation. Films and comics both require continuous storyboards, and brand advertising often also needs a whole set of visual materials. How the same character changes across different frames while remaining consistent, and how shots connect to one another—these problems cannot be solved merely by improving the image quality of a single image.
Kling Image 3.0 already supported continuous image sets, and with 4.0, Kling further strengthened subject and scene consistency across multiple images. According to the official introduction, the new model can generate up to ten interrelated images at once.
We selected a still from the film Portrait of a Lady on Fire as a reference and asked the model to generate a set of film storyboards on its own around the story of two women reuniting by the seaside and burning an unsent letter.

Still from the film Portrait of a Lady on Fire
From the results, the model was already able to organize a fairly complete visual narrative.

Excerpt of the 10 generated film storyboard images
The most striking aspect of this set of images is that the model proactively used different shot sizes—long shot, medium shot, close-up, and so on—advancing the visual narrative layer by layer. In particular, the transition from handing over the letter to the close-up of the envelope and then to the shift in the characters' expressions already shows a relatively clear awareness of shot organization.
The story begins with a long shot of two people on the coast, then switches to a medium two-shot in which one person hands an envelope to the other; immediately after that, the camera moves in to a close-up of the envelope; by the fourth image, the scene switches to a close-up of a character's face looking down at the letter in their hands; from the fifth image, the bonfire from the reference image appears, and it ends with a long shot of the two standing on either side of the firelight.
In addition, the visual style of this generated set of images is relatively unified. The dark clothing and firelight atmosphere from the reference still are carried through, and the gloomy coast, deep blue twilight, rocks, and warm yellow firelight run through the entire set of images.
Of course, a closer comparison still reveals some small problems. For example, the characters' faces change subtly across different shots, and the transitions between certain actions are not natural enough. However, for this test, which used a story synopsis as input, the model has already delivered a set of storyboard drafts with a clear narrative order and a unified visual style.

(A set of animation storyboards we asked Kling Image 4.0 to generate)
For early-stage film and television creation, comic storyboarding, or brand advertising planning, this capability offers a lot of imaginative potential. Creators can first provide the story, character settings, and visual references, let the model generate a set of preliminary frames, and then choose suitable shots from them for further adjustment.
In the real film and television industry, such AI image-generation tools are also entering pre-production work. In June this year, American director Martin Scorsese, who has been making films for decades, said in an interview with The New York Times that he had been drawing storyboards for about 70 years. One difficulty he has always faced is how to accurately convey the scenes in his mind to cinematographers, art directors, and other crew members. Now, he can first use image models to produce these images, allowing the team to understand and discuss them more quickly.
This is very similar to the scenario in our test this time: with only a story synopsis and one reference image, Kling Image 4.0 completed the first round of storyboard conception. It is more like a starting point that can be used for subsequent creative discussion. What Scorsese meant by "letting the team understand and discuss more quickly" is probably also this: the model first turns the images in the mind into something that can be pointed to, while the real judgment and revision remain in the hands of the creator.
After generation
When creating with AI, people often have two very different sets of requirements: when the idea is still vague, they hope AI will think more on their behalf and provide broader inspiration; when there is a clear design, such as a partial modification, they hope it will strictly follow instructions and not improvise.
These are actually two different capabilities: one is understanding the user's intent and completing the parts the person has not yet thought through; the other is accurately executing requirements and interfering as little as possible with a design that has already been finalized.
The technical upgrade of Kling Image 4.0 can be understood in this context. According to Kling's introduction, the new model adopts a unified autoregressive architecture for understanding and generation, and handles more complex creative requirements through instruction-processing capability of up to 15K tokens, structured visual grammar, and Agentic task-planning capability. What these technical terms ultimately aim to solve is precisely whether the model can understand a task, arrange its details, and preserve the parts the user is already satisfied with when modification is needed.
These changes are not happening only at Kling. In November this year, Adobe launched new creative Agents and design tools, attempting to put ideation, generation, and modification into the same working environment; Kling's Video 4.0 released in October also further emphasized multiple keyframes, reference materials, and shot control. Today's products have further added visual design and editing capabilities.
For Kling, there is also a more practical commercial consideration behind this. Kuaishou's financial report shows that in the second quarter of 2026, Kling AI's revenue exceeded 850 million yuan, a year-on-year increase of more than 200%. Video generation has already brought it considerable revenue, but as commercialization reaches this stage, new questions arise: how can users move from occasionally generating a video to continuously using Kling to complete creative work?
The upgrade of the image model provides more possibilities for this. What it is competing for is users' ongoing creative needs on the platform. A stunning image or video determines why users open Kling for the first time; whether it can support the entire process from inspiration to finished product may determine why they keep using Kling.
Click the keywords to view the corresponding hands-on test content
GLM 5.1 vs DeepSeek V4 Pro vs Kimi K2.6 vs Qwen 3.7 Max vs MiniMax M3
DeepSeek
