AI Image Models — Full Overview

First version — introduced the concept of text-to-image generation to the public. 12B parameter transformer model.

Major quality leap. Introduced inpainting and outpainting. 4× higher resolution than DALL-E 1. Used CLIP image embeddings.

Dramatically improved prompt adherence. Deeply integrated into ChatGPT. Strong text rendering in images.

First native image generation model integrated into GPT-4o. Best-in-class text rendering, strong instruction following.

Improved consistency, photorealism, and editing capabilities over GPT Image 1.

Major generational leap. Introduced Instant and Thinking variants — Thinking mode researches context before generating. Best-in-class text rendering, multilingual support, full magazine/comic layout generation. Replaces DALL-E 3 and GPT Image 1.x in the API (deprecated May 12, 2026).

Introduced by Google Brain. Outperformed DALL-E 2 on COCO FID at launch.

Substantially improved photorealism, text rendering, and multilingual support. Launched in Google Bard and Vertex AI.

Highest quality Imagen to date at launch. Improved detail, lighting, and artifact reduction.

Next-gen text-to-image with best-in-class landscape, architecture, and product photography. Integrated into Gemini.

Fast image generation + editing model based on Gemini 2.5 Flash. Integrated into Gemini apps.

Professional-grade image generation and editing. Precise instruction following, near-seamless inpainting.

First public release. Latent diffusion model — runs on consumer GPUs. Sparked the open-source AI art movement.

Improved training. Still the most widely used SD checkpoint — thousands of community fine-tunes built on it.

Higher resolution (768px default), new text encoder (OpenCLIP), new depth model.

Improved NSFW filtering approach. Less over-filtering than 2.0, better prompt adherence.

Major architecture jump — 3.5B parameters. Native 1024px output. Two-stage pipeline with refiner model.

Multimodal Diffusion Transformer (MMDiT) architecture. Major improvement in text rendering and composition.

2B parameter variant of SD 3. Lighter weight, runs on consumer hardware. Same MMDiT architecture as SD 3 but optimized for accessibility.

2.5B parameter model. Balanced between quality and speed — ideal for local use on consumer GPUs with less than 8 GB VRAM.

8B parameter model. Best quality in the SD 3.x family.

Distilled 4-step variant. Near SD 3.5 Large quality at a fraction of inference time.

One of the most downloaded SD 1.5 fine-tunes on Civitai. Optimized for photorealism, cinematic compositions, and anatomical accuracy — achieves high quality with minimal prompt engineering.

Initial alpha release via Discord.

Improved coherence and style.

Better quality, more artistic control.

New proprietary model architecture. Major quality jump. More coherent scenes.

Photorealistic quality leap. Detailed hands and faces. Introduced --ar and --style flags.

Improved default aesthetics, more accurate with simple prompts.

Sharpened images, improved aesthetics, introduced zoom-out feature.

Much more literal prompt following, improved text in images, better photorealism.

Refinement of v6 — improved image quality, better human anatomy, sharper details.

Major architecture change. New personalization system, significantly faster generation, improved realism.

Apache 2.0 open-source, 4-step distilled model. Fast generation, consumer-ready.

Guidance-distilled, 12B parameter Flow Matching model. Better quality than schnell. Non-commercial license.

API-only commercial model. Highest quality in FLUX.1 family.

Improved version of FLUX.1 pro — 6× faster, better prompt adherence and quality.

32B parameter model — generation, editing, and multi-reference image combining in one. Open weights for research and non-commercial use. 226k+ downloads on Hugging Face.

Commercial API tier. Production-grade generation and editing with 4MP output and multi-reference control.

Flexible tier between dev and pro — balances quality and cost for scalable production workflows.

Premium tier. Superior facial detail, enhanced text rendering, maximum photorealism.

Smallest and fastest FLUX models to date — sub-second inference on capable hardware. Available in 4B and 9B parameter versions. Open weights.

One of the first and most downloaded FLUX fine-tunes. Optimized for photorealistic output — improved natural color grading, skin detail, and landscape rendering. Fully open source.

Consistency distillation of FLUX.1 [dev] down to 8 sampling steps with near-identical quality. Makes local FLUX generation significantly faster on consumer hardware.

Face identity preservation for FLUX. Generate consistent portraits of a specific person without fine-tuning — just provide a reference photo. Popular for character consistency workflows.

Community-favorite portrait fine-tune. Natural skin textures, studio-quality lighting, and sharp hair detail. Widely regarded as the best FLUX fine-tune for commercial portrait work.

ControlNet implementation for FLUX — enables structural guidance via depth maps, canny edges, and pose skeletons. Brings the precise layout control from SD/SDXL to the FLUX architecture.

Adversarial distillation of FLUX.1 [dev] to just 8 steps with minimal quality loss. Developed by Alibaba's Alimama team — one of the fastest high-quality FLUX variants.

First commercially licensed AI image generator. Trained on Adobe Stock + Creative Commons. Integrated into Photoshop.

Improved photorealism and detail. Photo settings for lighting and depth of field.

Major quality improvement. Introduced Structure Reference and Style Reference features.

Advanced multi-entity generation, improved realism, better Creative Cloud integration.

Meta's first image generation model. Fine-tuned on curated high-quality data. Integrated into Instagram.

Instruction-based image editing model. Precise local and global edits via text commands.

Largest generative multimodal model from Meta. 37B parameters, strong few-shot visual generation.

Strong composition and color accuracy. Designed for professional and enterprise workflows.

Improved cinematic quality, better text support, enhanced style fidelity.

Open-source release. Built on enhanced SDXL. Recognized for vibrant colors and diverse skin tone rendering.

Image generation component of Kling. Consistent character generation and style fidelity.

Unified image and video architecture. Exceptional character consistency across outputs.

High-quality generation with East Asian aesthetic strengths. Available via Alibaba Cloud API.

Integrated into the Qwen multimodal model family. Strong product photography and cultural representations.

Novel S3-DiT (Scalable Single-Stream Diffusion Transformer). Efficient unified text+image processing.

First public release. Strong in portrait and lifestyle imagery.

Improved realism and cinematic composition.

Best-in-class portrait photorealism from Tencent. Warm color grading, studio-quality lighting.

Launch. Immediately recognized for best text rendering of any image model at the time.

Significantly improved photorealism while maintaining text advantages. New Style and Color Palette features.

Further photorealism improvements. Strongest text-in-image model available. New canvas editing features.

Microsoft Designer launched with DALL-E 3 backend. Integrated into Bing Image Creator and Edge.

Upgraded to GPT Image 1 backend. Integrated across Microsoft 365, Windows Copilot, and Designer.

Fast variant. 17B parameter model, 4-step distillation.

Development/research variant. Non-commercial license.

Full quality model. Exceptional skin tone accuracy, fashion and beauty photography.

xAI's first public image model, integrated into Grok on X (formerly Twitter). FLUX-based architecture. Notably less restrictive safety filters produce a hyper-realistic look, especially for people and social scenes.

Next-generation model. Continues the Aurora lineage with improved photorealism and native Grok-3 multimodal integration.

Early release focused on vector-style and design-oriented generation.

Improved style consistency, introduced SVG output support.

#1 on Hugging Face text-to-image leaderboard at launch. Industry-leading text rendering, logo generation, and vector graphic output. The go-to model for professional graphic designers.

Cinematic-style SDXL fine-tune. Strong in dramatic lighting, movie-like scenes, and character consistency.

Leonardo's first proprietary foundation model. Improved text rendering, prompt adherence, and photorealism. Millions of daily generations across the platform.

Refined flagship model post-Canva acquisition. Cinematic quality with better anatomy and enhanced style controls.

First public Apple image generation feature. Clean, illustrative style — Animation, Illustration, and Sketch modes. Runs fully on-device. Available on iPhone 15 Pro and later.

Sketch-to-image feature in Apple Notes. Turns rough hand-drawn sketches into polished illustrations. On-device, privacy-first.

AI-generated custom emoji from text descriptions. Integrated into keyboard and Messages.

Expanded style options and improved quality expected with iOS 19. Broader device support.

OpenArt's flagship model trained for maximum photorealism in portrait and product photography. Optimized for studio lighting simulation, natural skin rendering, and commercial-grade output.

Amazon's first image generation model on Bedrock. Designed for enterprise workflows — product visualization, marketing assets, content generation at scale. Features built-in watermarking.

Improved quality, new image conditioning features including background removal and outpainting. Better instruction following.

Launch model. Recognized for strong depth-of-field rendering and rich shadow detail.