MTEdit: How Packed Trajectory Data Is Revolutionizing Multi-Turn Image Editing Training

Multi-turn image editing is becoming an important direction for human-AI collaborative creation, but building training data has always been a challenge. Traditional serialized rollout methods accumulate errors at each round of editing, causing visual quality to drift round by round. The recently proposed MTEdit dataset offers a new solution to this dilemma through an innovative pipeline called "Packed Trajectory Canvas" (PTC).

The Training Data Dilemma of Multi-Turn Editing

Multi-turn image editing allows users to progressively modify an image through a series of instructions, which is closer to real creative scenarios than single-turn editing. However, obtaining high-quality multi-turn training data is not easy. Research points out that the conventional serialized rollout approach—where the model continues editing on top of the previous round's output—accumulates errors and introduces visual quality drift across rounds. Simply put, one wrong step leads to every subsequent step being wrong, and the image gets blurrier with each edit.

The root of this problem lies in the fact that each round of editing is built on a flawed preceding state, and errors snowball. For editing tasks that require content consistency, this drift is especially fatal.

Error Accumulation is a classic problem in sequential generation: when each round's output serves as the next round's input, small deviations in earlier steps are repeatedly amplified by later steps. In image editing, this deviation manifests not only as pixel-level blur or artifacts but also undermines semantic consistency—for example, if an object's color is changed incorrectly in one round, subsequent rounds may continue modifying on that basis, causing the final result to deviate from the original intent. Error accumulation in multi-turn editing is harder to correct than in text generation because images are continuous high-dimensional signals without a discrete-token "resampling" mechanism to interrupt error propagation. Therefore, avoiding accumulation at the data generation pipeline level is more effective than remedying it at the inference stage.

PTC: Generating Multi-Turn States in a Unified Visual Context

The core innovation of MTEdit is the Packed Trajectory Canvas (PTC) generation pipeline. Unlike round-by-round rollout, PTC generates multiple editing states simultaneously within a unified visual context, thereby reducing cross-round visual quality drift.

Think of it this way: rollout is like a relay race, where each leg can drop the baton; PTC is more like spreading multiple editing results across a single canvas that shares the same visual baseline. This joint generation strategy suppresses error accumulation from the source, allowing each round's editing state to maintain high visual quality.

The MTEdit dataset built on this pipeline covers two types of scenarios—stepwise editing and context-aware editing—and contains 500,000 multi-turn image editing trajectories, supporting up to 8 rounds, with an average of 5.38 rounds per trajectory. According to the original data, this is one of the largest multi-turn image editing datasets to date.

Joint Generation differs from autoregressive round-by-round rollout in its conditional dependency structure. In rollout, the state at round t depends on the actual output of round t-1, forming a chain dependency; joint generation, by contrast, conditions multiple editing states jointly on the same initial visual context (such as the original image or a unified canvas), so errors are not passed between rounds. This is similar to denoising multiple related samples simultaneously in a diffusion model while sharing the same noise baseline. The PTC "canvas" is essentially a spatial or latent-space container that packs multi-turn editing results together, enabling the model to observe global consistency constraints across rounds during training and thus learn to perform stepwise modifications while keeping content stable.

MTEdit-Bench: Evaluating Multi-Turn Editing Ability Across Four Dimensions

Data alone is not enough—evaluation criteria are equally critical. The research team also introduced MTEdit-Bench, which examines a model's multi-turn editing ability across four dimensions: following editing instructions, maintaining content consistency, resisting visual quality drift, and effectively leveraging cross-round historical context.

These four dimensions essentially cover the core challenges of multi-turn editing. In particular, "resisting visual quality drift" and "leveraging historical context" strike directly at the weaknesses of traditional methods. An excellent multi-turn editing model must not only understand each step's instruction but also remember what was changed before and avoid making things worse with each edit.

Content Consistency in multi-turn editing evaluation usually refers to the visual and semantic stability of non-edited regions or designated objects to be preserved. Unlike single-turn editing, consistency in multi-turn scenarios must be measured across all historical rounds: the model should remember which regions were required to be preserved, which have already been modified, and avoid mistakenly altering them in subsequent rounds. Common quantitative methods include computing perceptual similarity (such as LPIPS, SSIM) of regions outside the editing mask, and using vision-language models to judge whether object identity or layout has shifted. This dimension is complementary to "resisting visual quality drift"—the former focuses on whether the edits are correct, while the latter focuses on whether image quality has degraded.

Fine-Tuning Validation: Qwen-Image-MTEdit Tops Open-Source Models

To validate the effectiveness of the dataset, the research team fine-tuned Qwen-Image-Edit-Causal using MTEdit to obtain Qwen-Image-MTEdit. On MTEdit-Bench, they evaluated a wide range of image editing models, including both closed-source and open-source ones.

Experimental results show that Qwen-Image-MTEdit performs best among all evaluated open-source models, demonstrating the effectiveness of MTEdit in improving a model's multi-turn image editing ability. This result also indirectly indicates that the way training data is constructed—not just model architecture—has a significant impact on multi-turn editing performance.

Qwen-Image-Edit-Causal is a variant in the Qwen image editing model family oriented toward causal/autoregressive multi-turn editing. The choice to fine-tune on a causal base rather than a bidirectional diffusion model is because multi-turn editing naturally has sequential dependencies, and the causal structure is closer to the reasoning form of "generating the next round based on historical context." During fine-tuning, MTEdit's trajectory data provides supervision signals in the form of packed canvases, enabling the model to learn to jointly predict subsequent states given multi-turn history. The choice of this base also shows that improvements in the data pipeline can directly connect with existing autoregressive editing architectures without redesigning the model.

Implications for the Multi-Turn Image Editing Field

MTEdit's approach reveals an important trend: as model architectures gradually converge, the quality of data engineering often becomes the performance differentiator. The value of the PTC pipeline lies not only in expanding data scale but also in providing a systematic method to suppress error accumulation.

For developers and researchers focused on AI image editing, MTEdit's data construction paradigm is worth learning from—rather than repeatedly correcting errors in rollout, it is better to redesign the pipeline from the source of data generation. Of course, the method's generalization ability in more complex editing scenarios, as well as the actual training cost of 500,000 trajectories, still await further observation.

One-line Summary

MTEdit uses packed trajectory canvases to generate multi-turn editing data, suppressing error accumulation and setting new performance marks for open-source models.

Paragraph Summary

To address error accumulation and visual quality drift caused by serialized rollout in multi-turn image editing training data, MTEdit proposes a Packed Trajectory Canvas (PTC) pipeline that jointly generates multiple editing states within a unified visual context. The dataset built on this pipeline contains 500,000 trajectories, supports up to 8 rounds with an average of 5.38 rounds, and covers both stepwise editing and context-aware editing scenarios. The research team also introduced MTEdit-Bench, which evaluates models across four dimensions: instruction following, content consistency, resisting quality drift, and leveraging historical context. Qwen-Image-MTEdit, obtained by fine-tuning Qwen-Image-Edit-Causal, became the best-performing open-source model in the evaluation, showing that the way data is constructed has a significant impact on multi-turn editing performance.