Fable 5.2 Draws in MS Paint: An AI Art 'Regression' Experiment and Community Debate

Fable 5.2 paints pixel by pixel in MS Paint, sparking community discussion on AI capability boundaries and process transparency.
A Reddit post showed Fable 5.2 model painting pixel by pixel in Windows Paint. This "regression" demonstration under constrained conditions was compared by the community to rudimentary diffusion models, used to observe AI's planning ability without layers, anti-aliasing, and other professional tools. Discussion split into several threads: users called for demonstration videos to show clocks and token counts to quantify per-step time and computational cost; some mocked the model's capability as equivalent to a 7-year-old in the early 90s using a snowman output from Meta Muse Spark 1.3 free version; others suggested passing off AI artwork as human-made to test public bias. These seemingly scattered feedbacks collectively point to three core issues: AI painting's capability boundaries under constrained tools, lack of observability in generation processes, and identity perception controversy of AI art.
When AI Puts Down Its Brush and Picks Up MS Paint
A Reddit post showed a model called Fable 5.2 painting in Windows Paint (MS Paint). No Photoshop layers, no brush engine from professional drawing software—just the most basic pixel-level operations. This itself was a deliberate constraint experiment. The post quickly sparked discussion, with some comparing it to "rudimentary diffusion models," arguing that this clumsy pixel-by-pixel operation exposes certain fundamentals of current AI painting's underlying generation logic.
The value of this "regression"-style demonstration isn't in producing beautiful images, but in observing how the model plans under minimalist tool constraints. MS Paint has no advanced undo history management, no anti-aliasing; every brushstroke is an irreversible pixel stack, forcing the model to think like a human beginner about "what to draw first, what to draw next."

Diffusion Models are the core architecture of current mainstream image generation models. The basic idea: gradually add Gaussian noise to a real image until it becomes pure noise, then train a neural network to learn the "reverse" process—gradually denoising from noise to recover the image. During generation, the model starts from random noise and goes through tens to hundreds of denoising iterations to finally obtain a clear image. Each denoising step is performed in latent space or pixel space, and the model needs to predict and remove the current noise component at each step. The "rudimentary diffusion model" analogy in the article borrows this iterative "from nothing to something, gradually building" characteristic to describe the pixel-by-pixel painting process in MS Paint, although the latter doesn't involve true noise modeling but rather makes a pixel decision at each step.
What the Community Is Debating: Efficiency, Transparency, and Mockery of the "Free Version"
The discussion quickly split into several threads. The most direct technical feedback came from demands for process observability: one user wanted a clock in the bottom right corner of the video to know how long each step took; another immediately added "token count would be nice to see too." Both comments point to the same pain point—the AI painting process is currently a black box for onlookers, lacking quantifiable intermediate metrics to evaluate efficiency.
Another thread was ironic. A user posted the output of Meta Muse Spark 1.3 free version, showing a snowman, with a caption mocking its capability level. A replier added: "so that model is approx equal to a 7 year old in the early 90's." Though harsh, this comparison reflects a real market sentiment: users remain highly vigilant about the gap between capability claims and actual performance across models.
The most controversial comment had a social experiment flavor: suggesting posting the artwork in an amateur painting forum, claiming to hate AI and that this was hand-drawn, enjoying praise and comments like "AI could never do this," then revealing the truth. This comment resonated because it touches on the identity anxiety of AI art—when the value judgment of a work highly depends on the label of "whether it was made by human hands," the aesthetic evaluation of the work itself takes a back seat.
Token counting is a common unit of measurement in large language model (LLM) services: the model splits input text into tokens (roughly corresponding to a word or part of a word), and costs or measures based on the total tokens of input and output. In text generation scenarios, token count directly corresponds to computation and response length, serving as an intuitive efficiency metric. However, image generation models typically don't use the concept of token counting; their computational cost depends more on denoising steps, image resolution, and model parameter count. The community's demand to "see token count" essentially hopes to migrate the mature cost measurement approach from text generation to image/video generation demonstrations, thereby providing a comparable quantitative reference for the computational overhead of each step in AI painting.
Seeing AI Painting's Capability Boundaries Through Pixel Operations
Fable 5.2's performance in MS Paint actually provides a window into observing the model's spatial reasoning ability. AI features in professional drawing software usually appear as "generative fill" or "text-to-image," where the model only needs to output the final result; whereas painting pixel by pixel in MS Paint requires the model to have composition foresight similar to human artists—whether to lay down large color blocks first or outline first, how to handle pixel alignment, how to express light and shadow with limited tools.
One commenter described this process as "rudimentary diffusion models," and this analogy is worth pondering. The core of diffusion models is gradual denoising from noise, and pixel-by-pixel operations in MS Paint are likewise a gradual construction from nothing, except that the decision space at each step is compressed to a single pixel. Performance in such constrained environments may expose the model's weaknesses in long-term planning and local consistency more than fluent output in professional tools.
Spatial reasoning and composition foresight are among the core abilities of human artists: before putting brush to canvas, they establish the overall layout of the image in their mind, including subject position, proportional relationships, occlusion layers, and light direction, then implement step by step using a strategy of "overall first, then details" or "outline first, then details." For AI models, when directly outputting final results through "text-to-image" or "generative fill" in professional drawing software, the model doesn't need to explicitly plan each spatial operation; but in tools like MS Paint that only provide pixel-level brushes, the model must decide where the next pixel goes and whether the current local area is consistent with the overall composition. This actually requires the model to have serialized spatial decision-making ability—maintaining consistency between global composition and local pixels in the absence of advanced editing tools (layers, selections, undo history)—which is precisely where current models are relatively weak.
The Lack of Process Transparency: Calls for Clocks and Token Counts
The community's demand for "showing the clock" and "showing token count" is essentially a naive requirement for AI system interpretability. In video demonstrations, viewers can see the image changing but cannot judge the computational cost of each step. This information gap makes technical evaluation difficult—the same final image could be an efficient one-pass creation or the result of repeated trial and error, and viewers cannot tell.
This problem is especially prominent in AI painting. Text generation has token counting as a basic cost metric, but process metrics for image generation lack unified standards. Users' spontaneous demands actually point to a gap in product design: when AI painting is disseminated as demonstration content, visualization of process data should itself be part of the demonstration.
AI system interpretability refers to a series of technical and design practices that enable humans to understand, audit, and trust AI system decision processes. In generative AI demonstration scenarios, interpretability concerns not only internal model mechanisms but also external presentation: viewers need to know how much time each operation takes, how much computational resources are consumed, and whether repeated trial and error occurred. When demonstrations only show the final image or accelerated processes, viewers cannot distinguish between "an efficient single generation" and "screening after extensive trial and error." Therefore, the community's request to display clocks and token counts is essentially a demand for "process visualization," belonging to the end-user-facing transparency layer of interpretability design, complementing techniques like internal attention visualization and feature attribution.
Conclusion: A Small Experiment Reflecting Big Issues
Fable 5.2's MS Paint painting demonstration appears to be an entertaining technical showcase, but actually touches on three issues simultaneously: the model's true capability boundaries under constrained tools, the lack of observability in AI generation processes, and the identity controversy of AI art in public perception. Those seemingly scattered jokes and demands in community discussions—clocks, token counts, snowman comparisons, social experiment suggestions—when pieced together, form exactly a real user needs list. For developers and observers concerned with AI painting development, these voices are more worth reading carefully than any single performance evaluation.
Related articles

Zed Editor Releases v1.29.2: Observations on a Routine Version Iteration
Zed editor releases patch v1.29.2, version bump by @JosephTLyons, commit verified via GitHub signature. This article covers the release's basic info, what a patch version means, and practical upgrade advice for users.

Anthropic IPO Questioned: Hacker News Debates Timing of Going Public
A Hacker News post titled "Just say no to Anthropic's IPO" sparks debate on the timing of Anthropic's public listing and AI capitalization, touching on valuation bubbles and mission independence.

Hands-On with the Genie 4 World Model: Directing Worlds in Real Time Like a Puppeteer
Google DeepMind's Genie 4 world model in depth: Special Guest X's hands-on demo of real-time scene direction via prompts, custom interactive apps in Project Genie, and prompts-as-seeds reproducibility.