On April 22nd, GPT-image-2, which had been in limited testing just days before, officially launched its public beta. Its practical performance has sparked widespread discussion within the AI community. The most critical improvements over previous image generation models are: clearer text, posters that more closely resemble design drafts, and finally usable UI screenshots. This has prompted discussions about image generation models starting to be considered as production tools. First, let's examine the generated results:

Behind the higher granularity of the output lies a significant shift in the underlying technical approach. The dominant methodology in recent years has been based on the principles of diffusion models. The starting point is straightforward: if a clear image can be gradually transformed into random noise by adding 'snow,' then conversely, by progressively removing noise from that 'snow,' it should be possible to reconstruct an image. Thus, models are trained to perform one task: at different stages of noise, to determine "where the image should converge next." This method has been highly successful visually. It excels at handling continuous changes, such as lighting, textures, and character details. However, it faces an almost inherent structural limitation: generation happens almost "holistically," without a concept of sequence. In the process from noise to image, all elements emerge together. Characters, backgrounds, decorations, and text are all 'painted' out within the same convergence path. The model lacks the ability to "write the first character, then the second," because in its world, discrete units like "characters" do not exist. This explains why earlier models universally failed with text. When it sees "HELLO," it learns common stroke combinations; during generation, it produces a "text-like texture" in a certain area. Constraints like letter order, spelling rules, and sentence length are not part of its expressive system. Many teams attempted to compensate with more data and higher resolutions, but with limited success, as simulating discrete structures within a continuous system inevitably leads to errors at critical points. The change embodied by the GPT-image-2 generation of models occurs precisely at this juncture. It first changes how an image is represented. Through a visual tokenizer, the image is broken down into a series of discrete units, similar to tokens in text. This transforms the image into a sequence that can be generated step-by-step. Once in this sequence space, the mature methodologies of language models can be directly applied. The generation process gains a sequence; it can be "written out from start to finish." Order, length, and contextual constraints can all be explicitly controlled during this process. An even more critical step is the introduction of a training approach akin to an "agent." The characteristic of an agent is to first understand a task, then formulate a plan, and finally execute it. In GPT-image-2's generation pipeline, the language model assumes a role similar to a "planner." Based on the input, it decomposes the requirements into a structure—for example, identifying where the title goes, what content to write, its approximate position, and whether multi-line layout is needed. This process is invisible to the user but creates an implicit layout sketch within the model. Subsequently, the visual component performs the rendering constrained by this sketch. Text becomes a predefined objective. The order and content of characters are determined by the language model, while the visual model is responsible for presenting them in a suitable style. From an engineering perspective, this embeds a "planning-execution" pipeline within the model itself, giving it steps, structure, and intermediate decisions, much like an agent. The impact of this structure on text is immediate. This is because text is inherently a strongly constrained sequential task, and language models excel at handling sequences. Once these two are aligned, "writing the correct characters" no longer relies on luck but becomes a target that can be stably optimized. This is also why GPT-image-2 performs exceptionally well in scenarios like posters, UI, and e-commerce graphics. The difficulty in these scenarios has always been structure and constraints, not pure visuals. Once the structure is predetermined, the freedom of subsequent rendering becomes easier to control. Currently, most domestic models are at the intersection of these two paths. Doubao Image has begun incorporating language models into generation decisions, showing significant improvement with short Chinese text and simple layouts. This indicates that a "planning layer" is forming, but there are still fluctuations with long text and complex layouts, suggesting that the alignment between discrete representation and visual rendering is not yet stable. Kuaishou's Kolors is very impressive visually, with styles and textures approaching the industry's top tier. However, text is still largely compensated for in the visual stage, lacking prior constraints, making it prone to errors as text length increases. Alibaba's Qwen and Baidu possess advantages in data and scenarios, particularly within e-commerce and search ecosystems, giving them the conditions to build large-scale structured datasets. However, their current image generation primarily continues along the original path, with language models not yet serving as the core controller of the generation pipeline. From a methodological perspective, the gaps concentrate on three points: whether the image is discretized into units that can be processed sequentially, whether the language model is integrated into the main generation pipeline, and whether a data system with layout and text annotations has been established. Once these three elements are connected, text issues will essentially disappear. This path is also gradually converging with the development direction of text models. The core reason models like Claude are used by many developers for practical work is their greater stability in executing complex tasks. Capabilities like long-context handling, structured output, and step completeness make them more akin to systems that can deliver results. The evolution of the GPT series from conversational agents to tools is essentially about strengthening this "task completion" capability. Image generation is undergoing a similar phase: moving from "generating a visually pleasing image" to "completing a task with visual constraints." When language models, discrete representation, and agent-like planning mechanisms are combined, an image ceases to be merely a visual outcome and becomes a new medium for expression and execution.