AI Image Generation and Editing Models: Technical Capabilities, Evaluation, and Engineering Practices

    This article explains how AI image generation is shifting from one-shot creation to controllable, stable, and repeatable editing. It covers instruction-based editing, pixel-level preservation, native RGBA and layered output, photorealistic rendering, text rendering, quantitative evaluation such as GSB, billing models, cost control, MaaS, and cloud-based delivery.

    In recent years, image generation and editing models have undergone a major shift: from “being able to generate an image” to “being able to produce stable, controllable, and repeatedly editable results.” This shift places higher demands on developers. Beyond visual appeal, developers must consider controllability, editing stability, and engineering factors such as API cost and service stability. This article provides a systematic analysis of the key technical capabilities, evaluation methods, and engineering practices for image generation and editing models, helping developers build a complete technical framework.

    1. From Generation to Editing: A Shift in Technical Paradigms

    Traditional text-to-image models focus on generation from scratch: creating a complete image from a prompt in one pass. However, real-world scenarios in design, e-commerce, and content production often require local modification while preserving the overall image—changing a product background, adjusting clothing, or revising poster copy. This requires instruction-based editing.

    Compared with generation, editing faces two fundamental challenges:

    • Localization: understanding “what to change” and locating the target area based on natural-language instructions.
    • Preservation: modifying the target area while minimizing changes to the rest of the pixels, avoiding the problem of changing one part and disturbing the entire image.

    Whether a model can balance these two requirements is key to determining whether it is production-ready.

    2. Pixel-Level Preservation of Non-Editing Areas

    A core capability of high-quality editing models is pixel-level preservation of non-target areas, even after multiple editing rounds. Common technical approaches include:

    • Mask mechanism: using masks to separate editable and locked regions, and only repainting areas covered by the mask.
    • Conditional injection and feature fusion: injecting features of the original image as strong constraints during generation to suppress unwanted changes.
    • Attention constraints: constraining attention weights in cross-attention layers to prevent information from spreading to areas that should not be modified.

    A key engineering challenge is stability across multi-round editing. When users edit the same image dozens or even hundreds of times, errors can accumulate and cause drift: unrelated areas gradually degrade during repeated edits. Maintaining consistency of the subject, background, and composition across consecutive operations is an important stress test for the robustness of an editing model.

    3. Native Transparency (RGBA) and Layered Output

    For design and asset production, native transparent channel output is a highly practical capability. Traditional models usually output only opaque RGB images, and transparent backgrounds require post-processing, which is time-consuming and often inconsistent.

    Models with native transparency support can directly output RGBA images with an alpha channel and may even support layered output, separating foreground subjects, shadows, and backgrounds into independent layers. This can significantly reduce post-compositing costs in scenarios such as e-commerce product images, advertising assets, and UI icons. Developers can build asset pipelines that are ready to use immediately after generation, for example:

    1. Input: prompt + size + transparency (RGBA) + number of layers
    2. Output: PNG with alpha channel (optional: multi-layer PNG or PSD-style layered output)

    4. Photorealistic Performance: Lighting, Materials, and Textures

    The difficulty of photorealistic image generation lies in how faithfully the model reconstructs the physical world. High-quality models continuously improve in the following dimensions:

    • Lighting consistency: light source direction, shadow projection, reflection, and refraction follow physical rules.
    • Material reproduction: metals, glass, fabrics, and skin exhibit realistic roughness and highlight behavior.
    • Skin texture: close-ups of people retain details such as pores and texture, avoiding an overly smooth, plastic appearance.

    These capabilities directly determine whether an image is suitable for commercial photorealistic creation.

    5. Text and Glyph Rendering

    Posters, packaging, and infographics often require text inside images. Text rendering has traditionally been a weak point for generation models, often producing distorted strokes or missing characters. Two common metrics for evaluating text rendering quality are:

    • Severe text corruption rate: the proportion of text that is completely unreadable or structurally broken.
    • Minor text corruption rate: the proportion of text that is mostly readable but has small defects.

    Improving rendering of dense small text and complex glyphs—such as multi-stroke characters and artistic fonts—is an important direction for evaluating a model’s text capabilities. In engineering practice, if a model’s text capability is insufficient, a hybrid approach of generating a base image and overlaying vector text is often needed to ensure accurate copy.

    6. Evaluation Systems: How to Quantify Image Generation Quality

    Evaluating an image generation or editing model should not rely solely on subjective impressions. A multi-dimensional quantitative system is needed:

    1. Objective metrics: automated measures such as text corruption rate, structural consistency, and image fidelity.
    2. Subjective evaluation: human ratings of aesthetics and instruction following.
    3. GSB (Good/Same/Bad) comparative evaluation: comparing a new model with a baseline model on the same inputs and measuring the proportion of cases that are “better / same / worse.” This produces relative advantage values such as “+55.62%.” The advantage of GSB is that it closely reflects real user preference and is suitable for subjective tasks such as image-to-image generation.

    For developers, it is recommended to build an internal evaluation set that covers typical business scenarios and forms a regression quality baseline, avoiding unquantifiable judgments such as “it feels better.”

    7. Billing Models and Engineering Cost Control

    Commercial deployment of models requires reasonable billing design. Image generation services commonly use two billing approaches:

    • Two-stage billing (input image + output image): editing tasks are billed based on the size of the actual input image and output image.
    • Text token billing: billing based on the amount of input and output text.

    A developer-friendly billing design often meters editing tasks by “input image + output image,” while charging little or nothing for the text portion, keeping the cost of a single edit predictable. When selecting a model, developers should focus on:

    • Cost per image for each editing or generation task.
    • Whether concurrency throughput meets peak business demand.
    • Whether batch and asynchronous interfaces are available to reduce per-call overhead.

    On the engineering side, costs can be further optimized through strategies such as result caching, resolution tiering, and on-demand calls.

    8. Model as a Service (MaaS) and Ecosystem Collaboration

    Image generation capabilities ultimately need to be delivered to developers through platform-based services. Model as a Service (MaaS) platforms usually take on the following responsibilities:

    • Unified access: hiding differences between underlying models through a standard API and providing a one-stop calling experience.
    • Compute adaptation: adapting and optimizing models for different heterogeneous compute architectures to improve cost-effectiveness.
    • Token metering and settlement: establishing metering, quotas, and settlement systems for model calls to support token-market-style resource circulation.
    • Scenario packaging: packaging general model capabilities into industry-specific solutions.

    At the ecosystem level, platformization creates value in several directions:

    • Intelligent product matrix: combining image generation, text generation, and code generation into end-to-end solutions.
    • AI Coding integration: combining code generation with business workflows to improve development efficiency.
    • Industry-integrated solutions: providing “compute + model + application” delivery for industries such as finance and government, which have higher requirements for compliance and stability.

    One-stop AI cloud platforms, such as Baidu AI Cloud, typically provide general cloud capabilities such as container platforms, model services, object storage, logging services, and monitoring and alerting. These can help developers complete the full cycle from model access and elastic scaling to observable operations without building underlying infrastructure.

    Conclusion

    The technical competition among image generation and editing models has expanded from a single focus on “generation quality” to a comprehensive comparison of “editing controllability, photorealistic expression, text rendering, billing cost, and platform delivery.” For developers, mastering these technical dimensions and evaluation methods enables rational model selection. At the same time, mature MaaS platforms and general cloud capabilities can significantly lower engineering barriers, allowing developers to focus on business innovation.