How AIGC Is Revolutionizing Image Generation and Editing

Introduction

With the rapid development of artificial intelligence, generative AI (AIGC) is profoundly changing enterprise operations and marketing, especially in the image domain, driving a shift from "generative" to "controllable generation".

Speaker

Cheng Bo, senior algorithm expert at 360 AI Research Institute, graduated from Xi'an Jiaotong University and previously worked at Baidu, Didi, and Tencent, brings deep expertise in video and image recognition, understanding, and AIGC.

Outline

AIGC and the "new revolution" in the image field

Frontier exploration of generation and editing

Business applications and capability considerations

Future outlook

1. AIGC and the New Revolution in Image Domain

Traditional image creation suffers from high labor cost, low efficiency, and creative limitations. AIGC offers three core advantages: second‑level generation (seconds per image), zero‑threshold creation (non‑experts can use), and cross‑style fusion (style‑guided generation).

Since 2024, mainstream large models can generate images in under one second, and image‑related AI applications now account for about 20% of AI consumer‑type revenue, surpassing language‑only tools.

2. Core Concepts of AIGC

AIGC (Artificial Intelligence Generated Content) in the image field consists of two main branches: Image Generation (text‑to‑image) and Image Editing (image‑to‑image). The technology has evolved from GANs to Diffusion Models, Flow Models, and multimodal interaction.

3. Key Technical Modules

Multimodal Input Processing

This module encodes various inputs (text, images, sketches, audio, etc.) into a unified representation that the model can understand, e.g., using CLIP‑style alignment.

Image Editing Core Capabilities

Local editing: intelligent inpainting, background removal, and region replacement.

Outpainting: extending image borders while preserving style.

Attribute modification: precise changes to color, material, or other properties.

Style transfer: extracting a reference style and applying it to a target image.

Model Optimization Techniques

Quantization compression: reducing parameter precision to lower compute and memory usage.

LoRA fine‑tuning: injecting low‑rank adapters for efficient task‑specific adaptation.

Distributed inference: parallelizing computation across multiple devices for faster large‑image generation.

4. Controllable Generation Techniques

Structure‑Constrained Generation

Uses control signals such as sketches, line drawings, depth maps, or layout maps to define the spatial framework of the output.

Semantic‑Constrained Generation

Leverages detailed textual triples, object attributes, and scene descriptions to guide content and relationships (e.g., FLUX, Qwen‑Image).

Style‑Constrained Generation

Employs reference style images, artistic tags, or color palettes to dictate the overall visual tone (e.g., Stable Diffusion Style Reference).

5. Controllable Editing Techniques

Local‑Constraint Editing

Inpainting with masks to replace specific regions while preserving surrounding context.

Local detail enhancement for faces, hands, or textures without altering overall composition.

Parameter‑Constraint Editing

Style intensity sliders to adjust the strength of transferred styles.

Bounding‑box manipulation for object size and position adjustments.

Interaction‑Constraint Editing

Brush‑driven real‑time editing where user strokes become edit commands.

Drag‑based pose adjustment (e.g., DragDiffusion) for intuitive character manipulation.

6. Representative Research Results

HiCo: layout‑controllable generation model built on Stable Diffusion, using multi‑branch architecture for precise multi‑object placement.

Qihoo‑T2X: efficient T2X model supporting text‑to‑image, text‑to‑video, and multi‑view generation, employing sparse attention to reduce complexity.

BDM: native Chinese image generation model with a dedicated language branch to improve cultural understanding.

PlanGen: multimodal understanding and generation model using a unified autoregressive framework for layout‑image joint generation and editing.

NAMI: high‑performance model with parallel slicing across time, space, and model dimensions, achieving a 64.8% speedup at 1024 resolution.

7. Business Applications and Capability Considerations

In e‑commerce, AI can automatically generate rich product scene images, virtual try‑on, and text super‑resolution, dramatically cutting cost and cycle time. In creative design, AI enables rapid multi‑style poster creation, layer‑aware generation for complex compositions, and entertainment‑focused style conversion (e.g., instant “abs‑muscle” photos).

8. Self‑Developed vs Open‑Source Solutions

Self‑developed solutions offer high customization, data security, and the potential to build innovation barriers, but they require high R&D cost, steep technical thresholds, and slower iteration. Open‑source solutions provide low development cost, fast deployment, rich community resources, and rapid updates, yet they may lack deep customization and pose security, IP, and compliance risks.

9. Future Outlook

Technical challenges include generation quality instability, high computational cost, and limited generalization for niche styles. Ethical and compliance risks involve copyright disputes, deep‑fake misuse, and bias amplification. Future directions focus on pixel‑level controllability, real‑time high‑resolution generation on consumer devices, and natural‑language driven interactive editing.

Overall, AIGC image technology will become more ubiquitous, powerful, and user‑friendly, driven by faster generation, lower barriers, and broader applicability.