In late October, inference platform fal.ai published a hands-on tutorial on making a video ad with Kling 4.0. One detail caught my attention. Kling generated the entire video, but the reference stills for every product and character were first made in Nano Banana 2.1, then brought into Kling to keep the subjects consistent.
The Kling 4.0 video model, released on October 15, could already generate 30-second clips directly and accept 10 keyframes and up to 15 multimodal references. Yet the previous image model, Kling Image 3.0 Omni, had fallen to No. 87 on the Artificial Analysis text-to-image leaderboard with an Elo of 925. As the video tools got better, the difficulty of getting a clean first frame became harder to ignore.
Thirty-three days later, on the evening of November 17, Kling launched Kling Image 4.0 on the web and through its developer API at the same time.
By late 2026, the AI video world already had plenty of good-looking one-off images.
On real projects, three kinds of rework eat up the hours:
- Character details drift when a still becomes the first frame of a video.
- A set of ten visuals gets less consistent with each new image.
- Posters and product pages with dense copy still have to go back into an image editor for typesetting and upscaling.
Kling Image 4.0 is an effort to handle all three earlier, at the image generation stage.
Three task modes across two API endpoints
It is easy to read the launch materials and assume “native 8K,” “10-image sets,” and “long product detail pages” all refer to the same capability. The official Kling Image 4.0 API usage guide puts them across two endpoints and three task modes, with very different limits.
An API endpoint is the specific address that handles a request. Think of the API as a service center with separate counters for different tasks. Although the product name is Kling Image 4.0 in both cases, developers send requests to two different addresses:
- Standard images and series:
POST /image-generation/kling-4.0(2K, 4K, or 8K, with up to 10 reference assets).- Long images:
POST /long-image-generation/kling-4.0(for e-commerce detail pages and long webpage visuals, with a 2K ceiling and up to three reference images).
| Task mode | Endpoint and key parameter | Output per request | Supported resolutions | Maximum reference assets | Maximum input length |
|---|---|---|---|---|---|
| Standard image or multiple views in one image |
POST /image-generation/kling-4.0, result_type="single"
|
1–9 independent images (default 1) | 2K / 4K / 8K (default 2K) | Up to 10 reference images and subject elements combined | 50,000 characters for English-only input or 21,000 for non-English-only input |
| Consistent image series |
POST /image-generation/kling-4.0, result_type="series"
|
2–10 images with a consistent visual identity (default 4, supports auto) |
2K / 4K / 8K (default 2K) | Up to 10 reference images and subject elements combined | 50,000 characters for English-only input or 21,000 for non-English-only input |
| E-commerce detail page or long webpage visual |
POST /long-image-generation/kling-4.0, skill="ecommerce" or skill="webpage"
|
One complete vertical image with multiple sections | 1K / 2K only (default 1K) | Up to 3 reference images. E-commerce requires 1–3. Subject elements are unsupported. | 166,000 characters for English-only input or 71,000 for non-English-only input |
Here is how the table maps to common jobs:
- To put front, side, and rear views of a character on a single white-background reference sheet, use
single. - To generate a six-frame storyboard or a coordinated set of main and supporting product shots, switch to
series. - To make a whole product page as one scrolling image, use the dedicated long-image endpoint, which tops out at 2K.
Underneath all three modes is what Kling calls its Native Unified Understanding and Generation Architecture. According to the official launch post, Kling Image 4.0 handles visual understanding and generation in a single autoregressive framework, with a prompt window of up to 15,000 tokens. At the API validation layer, standard image requests accept up to 50,000 English characters or 21,000 non-English characters, while the dedicated long-image endpoint raises those limits to 166,000 and 71,000 respectively. Mixed-language prompts count proportionally.
Editorial callout based on Kling AI’s launch post. The model handles visual understanding and image generation in one autoregressive framework.
The input window supports up to 15,000 tokens
Early diffusion-based image pipelines often passed a product brief of several thousand characters through a prompt rewriter that reduced it to a few hundred words of visual tags. Layout hierarchy and connections between images could disappear in that compression. With understanding and generation in one autoregressive sequence, the model can process the whole outline, plan where the headline goes, decide which sections explain the selling points, or work out how shots connect before rendering the pixels.
Kling has yet to disclose the model size, the structure of its autoregressive visual vocabulary, or per-request inference time. The public material establishes the architecture and the behavior of the interfaces.
A brief of up to 15,000 tokens and up to 10 reference assets assigned with @id enter the same autoregressive sequence. Outputs split between the native 8K standard and series endpoint and a separate long-image endpoint limited to 1K or 2K.
Native 8K, distinct faces, and roles for 10 assets fill the gap before video generation
Seen in the context of Kling’s own release schedule, the most urgent job for Kling Image 4.0 is to bring back the video preproduction work that had gone to other image models.
The first piece is native 8K output. In commercial design and film storyboarding, the reason to care about high resolution is usually the crop that comes later. A native 8K wide character reference can be reframed into a close-up or printed for a large backlit display, with detail still visible in hair, fabric, and skin. Upscaling a 1K or 2K source can leave fine details looking smeared or invent texture that was never there. Official launch post
Native 8K gives high-resolution displays and later crops more room
The model includes more than 110 lighting and tonal presets. A reference image can also supply its palette and tonal range, saving a separate LUT pass after generation.
English-labeled editorial recreation based on Kling’s published native 8K portrait and crop example. Kling’s source example shows hair, skin texture, and leather stitching without a separate upscaling pass.
The second piece is an effort to give different people in the same frame distinct faces. In group portraits, family photos, and street scenes with several characters, image models have often reused the same facial template. Kling’s international launch release and API guide single out improvements in multi-person scenes. Facial structure, build, age, and pose can vary across people in the frame. That is useful for defining a short drama’s ensemble cast or making an ad with several characters.
The third piece is assigning different jobs to different reference assets. Both single and series accept up to 10 assets in one request. Those assets can be individual reference images (refer_image) or multi-angle character or product packs already stored in Kling’s subject library (element). The prompt can assign each one a job using @id.
For example, use subject 1 to lock the lead character’s face, image 2 for the cut of the trench coat, image 3 for the material of the handbag, and image 4 for the interior lighting and spatial layout.
Each reference image can be up to 30 MB in JPG, JPEG, or PNG, with both sides at least 300 px and an aspect ratio between 2:5 and 5:2. There are nine output options: auto, 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, and 21:9. Custom ratios are unsupported. API usage guide
The design fits the Kling 4.0 video API opened in late October, which accepts up to 10 keyframes and 15 multimodal Omni References. With the same account and subject library, an artist can now use single to make a character turnaround sheet, save it as an element, and use series to lay out storyboard stills that feed directly into the video timeline. That closes the gap that previously sent first-frame work to a third-party model.
2,000 characters in one image and 10 consistent images per set: what rework goes away?
Native 8K and multiple references are aimed at filmmaking and commercial photography. The 2,000-character limit, consistent image series, and long images tackle the layout work behind e-commerce visuals, social image sets, and presentations.
For text rendering, Kling Image 4.0 raises its stated maximum to 2,000 characters in a single image. It covers seven languages: simplified Chinese, traditional Chinese, English, Japanese, Korean, Spanish, and Portuguese. It also includes more than 70 preset commercial-use fonts. A prompt can specify a font hierarchy and a font name, such as Font: Nova Grid in the guide. Font reference images can guide the letterforms as well. The goal is to lay out a whole menu, event schedule poster, infographic, or bilingual card in one pass. Official launch post · API usage guide
The API guide’s example of a named font, Nova Grid
English-localized illustration of Kling’s published text-rendering and editing examples. The upper-right panel recreates the original Chinese calligraphy as “Open Country.” The editorial recreations illustrate the source examples.
For multi-image and long-image layouts, the model brings work that used to require several manual requests into a single prompt.
Consistent image series (series). Given a story outline, a product launch brief, or a presentation outline, the model plans what each image should do and produces 2–10 images with consistent colors, fonts, subjects, and layouts. This works for storyboards, picture book spreads, social media carousels, and presentation visuals. With the count set to auto, Kling decides how many images to make, and the actual output count determines how much concurrency the request consumes.
E-commerce and webpage long images (long-image). The e-commerce mode (skill="ecommerce") requires one to three product images to anchor the appearance. It has three lengths: short (about 2–5 modules), medium (the default, about 6–10), and long (about 11–18). It can lay out selling points, specs, and use cases as one product detail image. The webpage mode (skill="webpage") generates long visuals for brand sites and campaign landing pages, accepts up to three optional references, and has no length setting. API usage guide
The example below starts with the reference image and prompt, followed by the resulting long product detail image. See “6 Featured Capabilities and Usage” in the API guide for more examples.
Asset roles: @ref_1: red electric hot pot product.
Generate an English e-commerce product page for a pot
English-localized editorial illustration based on the first four pages of Kling’s public series example: the cover, a philosophy spread, a 1996–2022 timeline, and a workshop close-up. The example uses a consistent K mark, cool concrete tones, and pagination.
The first layout takes less work, but three differences between the browser studio and the developer API matter in production.
Editing the text after generation is a browser feature. The launch post shows changes to the rendered text and font styles, along with panels for browsing 70-plus commercial-use fonts and 110-plus tonal presets. Those controls belong to Kling’s web-based Creative Studio. The long-image API returns a single flattened bitmap. The guide specifies that it provides no separate illustrations, HTML code, or editable design source files.
Dense small text still needs proofreading. Fitting 2,000 characters into an image does not guarantee that every character will render correctly. Section 4.2 of the API guide points to language, font complexity, text density, and layout as factors. Names, numeric specs, punctuation, and small dense text need a human pass before delivery.
- The long-image endpoint has its own limits on resolution and references. It tops out at 2K and defaults to 1K. It accepts up to three reference images and has no support for multi-image subject elements. This is useful for a quick product page direction or a first pass at a campaign visual. A responsive page that people can click through, or a design file whose product layers need to be edited separately, still has to be finished in Figma or front-end code.
Rankings and costs: the top Chinese model on these boards, with three numbers to unpack
As of November 18, 2026, independent evaluator Artificial Analysis listed Kling Image 4.0 on both its text-to-image leaderboard (AA-Image-T2I v2.1) and image editing leaderboard (AA-Image-Editing v2.1).
| Model | AA text-to-image rank and Elo | Default text-to-image price (USD per 1,000) | AA image editing rank and Elo | Image editing price (USD per 1,000) |
|---|---|---|---|---|
| OpenAI GPT Image 2.6 Sunburst (max) | No. 1 (1261 ± 14) | 210.72 | No. 2 (1211 ± 13) | 210.72 |
| Microsoft MAI-Image-2.7 | No. 4 (1214 ± 14) | 34.82 | No. 1 (1215 ± 14) | 34.82 |
| Kuaishou Kling Image 4.0 | No. 7 (1180 ± 13) | 63.00 | No. 6 (1183 ± 14) | 66.00 |
| Google Nano Banana 2.1 | No. 8 (1176 ± 12) | 33.60 | No. 10 (1152 ± 12) | 33.60 |
| ByteDance Seedream 5.5 | No. 14 (1134 ± 12) | 20.00 | No. 8 (1160 ± 14) | 28.00 |
| Black Forest Labs FLUX 3 Image | No. 21 (1115 ± 12) | 48.00 | No. 22 (1119 ± 13) | 48.00 |
| Alibaba Qwen-Image-3.0-Pro | No. 24 (1094 ± 12) | 40.00 | No. 29 (1092 ± 12) | 40.00 |
| Kuaishou Kling Image 3.0 Omni (previous generation) | No. 87 (925 ± 15) | 28.00 | No. 68 (954 ± 13) | 28.00 |
Text-to-image rankings as of November 18, 2026
Image editing rankings as of November 18, 2026
Three things stand out in that table.
-
Where Kling ranks and what another benchmark says. Compared with Kling 3.0 Omni, which launched in February, Kling Image 4.0 gained 255 Elo points on the text-to-image board, where its 95% rank interval was 6–9 across 5,550 comparisons. It placed sixth in image editing, with a rank interval of 5–7 across 5,628 comparisons. The official release also cites a No. 3 global result for its preview version in the UI/UX design category. Only OpenAI’s model family, with multiple GPT Image 2.6 and 2.5 variants, and Microsoft’s MAI-Image-2.7 are ahead. That puts Kling first among Chinese image models on these boards and just ahead of Nano Banana 2.1, which creators had been using to make first frames.
Change the evaluation criteria and the result changes. On DesignArena, a crowdsourced leaderboard that leans more heavily on blind judgments of visual design, Kling Image 4.0 is No. 15, with a 61% overall win rate. Image editing is its strongest category there as well. It scores well on instruction following, consistency when editing, and layout structure, while the design-focused blind test gives a less dominant result.
How the billing changed. The table shows a default text-to-image price of $63 per 1,000 images, or $0.063 each, in line with fal.ai’s launch pricing. The image editing board, which uses one reference image, shows $66 per 1,000. The November 17 official API changelog explains the underlying change: older Kling image endpoints charged for outputs, while Kling Image 4.0 charges for both input reference images and output images, and higher output resolutions cost more. Kling’s pricing page dynamically converts usage into package credits. The static documentation gives no separate dollar prices for 4K, 8K, or long-image outputs. A request with 10 reference assets and an 8K series can cost substantially more than the base price of $0.063.
The cost comparison and how long files stay available. At $63 per 1,000 images, Kling’s default price is under 30% of OpenAI GPT Image 2.6’s $210.72. Microsoft’s MAI-Image-2.7, however, scores more than 30 points higher on both boards and has a default price of $34.82. ByteDance’s Seedream 5.5 is still in the $20–28 range. For production, a less expensive approach is to validate composition, storyboards, and text layout at the default 2K resolution with a short reference list, then switch to 4K or 8K for the selected output.
One operational detail deserves a place in the cost calculation. The API returns links with hotlink protection and a 30-day retention window. Copy the files to your own object storage when the results come back. API usage guide
English recreation of the November 18, 2026 snapshot. Artificial Analysis text-to-image v2.1 Elo and default prices per 1,000 images, with notes on billing, web studio and API outputs, and retention.
Which mode to use depends on the deliverable
With the architecture, API limits, and bills in view, the choice comes down to what has to ship.
-
Kling video ads, character references for short dramas, and multi-shot storyboards. Bring the stills work into Kling Image 4.0. Use
singleorseriesand assign assets with@idto keep characters, wardrobe, and sets consistent. Start at 2K for routine character references, then go to 4K or 8K when the shot will need a tight crop or a push-in. This cuts out the visual mismatches that can come from importing images made on other platforms. - Posters with long Chinese copy, social visuals, and early e-commerce detail pages. Start with Kling’s web-based Creative Studio. It offers more than 70 commercial-use fonts, more than 110 tonal presets, and in-place text editing after generation. If the long-image API will make detail pages in bulk, plan for a 2K ceiling and a single flattened bitmap. Check names and numerical specs by hand before publishing.
- A general-purpose text-to-image or image editing API. If the work depends on multilingual layouts with 2,000 characters, native planning for sets of up to 10 images, or native 8K, Kling Image 4.0 is one of the few options that brings all three into the same API. For routine single-image generation and editing, compare Microsoft’s higher-scoring, nearly half-priced MAI-Image-2.7. For high-volume, low-cost assets, look at Seedream 5.5.
-
A responsive webpage or a layered design that will go through several revisions. Use
long-imageto agree on the visual direction quickly, then finish in front-end code or Figma. Treat the flattened visual as a tool for exploring the design. Handoff still needs an editable source.

















Top comments (0)