Probing Visual Planning in Image Editing Models
| Authors | Zhimu Zhou et al. |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2604.22868 |
| Download | |
| Code | https://github.com/spatigen/amaze |
Abstract
Visual planning represents a crucial facet of human intelligence, especially in tasks that require complex spatial reasoning and navigation. Yet, in machine learning, this inherently visual problem is often tackled through a verbal-centric lens. While recent research demonstrates the promise of fully visual approaches, they suffer from significant computational inefficiency due to the step-by-step planning-by-generation paradigm. In this work, we present EAR, an editing-as-reasoning paradigm that reformulates visual planning as a single-step image transformation. To isolate intrinsic reasoning from visual recognition, we employ abstract puzzles as probing tasks and introduce AMAZE, a procedurally generated dataset that features the classical Maze and Queen problems, covering distinct, complementary forms of visual planning. The abstract nature of AMAZE also facilitates automatic evaluation of autoregressive and diffusion-based models in terms of both pixel-wise fidelity and logical validity. We assess leading proprietary and open-source editing models. The results show that they all struggle in the zero-shot setting, finetuning on basic scales enables remarkable generalization to larger in-domain scales and out-of-domain scales and geometries. However, our best model that runs on high-end hardware fails to match the zero-shot efficiency of human solvers, highlighting a persistent gap in neural visual reasoning.
Engineering Breakdown
Plain English
This paper challenges the standard approach to visual planning in AI by proposing EAR (editing-as-reasoning), which reformulates visual planning as a single-step image transformation instead of iterative generation steps. The authors argue that traditional planning-by-generation methods are computationally expensive, and they test their hypothesis using abstract reasoning tasks like mazes and the Queen problem through a new procedurally generated dataset called AMAZE. The key insight is that visual reasoning doesn't need to be solved verbally or step-by-step; instead, a single intelligent image transformation can encode complex spatial reasoning directly. This approach aims to be both more efficient and more aligned with how humans visually plan compared to existing methods.
Core Technical Contribution
The core innovation is the editing-as-reasoning paradigm that reframes visual planning as a deterministic single-step image transformation rather than autoregressive step-by-step generation. Instead of a model predicting intermediate planning states sequentially (which is computationally expensive), EAR directly transforms a problem state into a solved or partially solved state in one forward pass. The authors introduce AMAZE, a procedurally generated benchmark with abstract spatial reasoning tasks (Maze and Queen problems) specifically designed to isolate reasoning capability from visual recognition, allowing researchers to measure pure planning ability without confounding visual understanding. This decoupling of reasoning from recognition is novel—previous visual planning work either conflated these or relied on language-based reasoning, whereas this work shows visual planning can be direct and single-step.
How It Works
The method takes an input image representing a visual planning problem (e.g., a maze with a start and end point, or a chessboard with a Queen placement puzzle) and passes it through a learned transformation function that outputs a modified or solved version of that problem in a single step. Rather than the model generating intermediate reasoning steps or using language as an intermediate representation, the image-to-image transformation directly encodes the reasoning process into pixel space. The transformation is learned end-to-end on the AMAZE dataset, which procedurally generates diverse problem instances with known solutions, allowing for exact supervision. The architecture likely uses a vision transformer or CNN-based encoder-decoder structure that learns to propagate and manipulate spatial information within the image itself. By framing planning as image editing, the model bypasses the computational overhead of sequential generation—no need for iterative decoding or language tokenization—and instead leverages the efficiency of single-pass feedforward computation.
Production Impact
Adopting this editing-as-reasoning approach could significantly reduce latency and memory requirements for visual planning tasks in production systems. Instead of auto-regressive models that require multiple forward passes (e.g., 10-50 steps for traditional planning-by-generation), EAR performs a single forward pass, reducing inference time from hundreds of milliseconds to tens of milliseconds depending on model size. For applications like robotics path planning, autonomous vehicle navigation, or interactive image editing tools, this speed improvement is substantial—real-time planning becomes feasible on edge devices. However, the trade-off is that the model must be trained on paired (problem, solution) datasets that cover the distribution of problems you care about; if your planning tasks are out-of-distribution from training, single-step reasoning may fail silently. Integration would require reframing your planning problems as image-to-image tasks, which may not be natural for all domains (e.g., constraint satisfaction problems that don't have clear visual structure).
Limitations and When Not to Use This
The approach is fundamentally limited to planning problems that can be naturally represented and solved in image/pixel space; abstract reasoning that lacks spatial structure or requires explicit logic reasoning may not benefit from this method. The paper uses relatively simple abstract puzzles (mazes, Queen problems) as evaluation tasks; it's unclear how well single-step visual planning scales to complex real-world problems like multi-step navigation in photorealistic environments or planning with high-dimensional constraints. The method assumes that the relationship between problem state and solution can be learned from examples, which requires large amounts of paired problem-solution data; for novel planning domains, collecting or generating this data could be expensive. Additionally, the approach may struggle with open-ended problems where multiple valid solutions exist or where the solution requires fundamental domain knowledge not present in the visual representation—editing-based reasoning cannot invent external knowledge the way language-grounded approaches can.
Research Context
This work builds on recent progress in visual reasoning and image-to-image translation models (like diffusion-based image editing), but applies this machinery to the neglected problem of visual planning. It contrasts with the dominant paradigm of language-grounded planning (e.g., using LLMs for chain-of-thought reasoning) and with prior visual planning work that relied on step-by-step generation. The introduction of AMAZE as a benchmark fills a gap in visual reasoning evaluation—existing datasets focus on recognition or object detection rather than abstract spatial reasoning. This research opens a new direction for visual reasoning: the hypothesis that complex planning can be learned as learned visual transformations, potentially inspiring follow-up work on what inductive biases (architectural constraints, training procedures) best capture planning in visual form.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::