Do You Remember? - That Summer on Port 7860

A Memoir of AUTOMATIC1111 Stable Diffusion WebUI

From a single line of Gradio script in August 2022, to the closing curtain of an era.

2022.08.22 - 2026.07.10 | An open-source legend of roughly four years

Prologue - Do You Remember?

Do you remember? It feels like it just happened yesterday. An anonymous GitHub user, a Gradio script found in a post on an anonymous image board, a git clone, a run.bat, and then a page popped up in the browser at 127.0.0.1:7860 - prompt on the left, generated image on the right, the progress bar inching forward step by step, VRAM usage jumping around in Task Manager.

That thrill of getting Stable Diffusion running on your own machine for the first time, that sense of accomplishment when you finally produced a "decent-looking" image after tweaking parameters for hours - thinking back on it now, it all still feels vivid.

But do the math: from that afternoon in August 2022 when the repository was created, to this midsummer of 2026 - it's been nearly four years.

Four years. On the timescale of internet products, four years is enough for a platform to be born, peak, and fade into obscurity; enough for a generation of users to dive in, drift away, and forget. A1111 WebUI traced this complete arc.

Four years ago, AI image generation was, for most people, synonymous with cloud services like Midjourney - the experience of typing /imagine in Discord and waiting for results. Running Stable Diffusion on a local machine meant opening a terminal, typing command-line instructions, navigating Python dependency hell, and facing a pile of incomprehensible parameters. A1111 packed all of that into a web interface, letting anyone who could type generate images on their own GPU. It wasn't the first to do this, but it was the most complete, the most user-friendly, and the one with the most thriving ecosystem.

And now? Open the GitHub Release page, and the last version sits at v1.10.1, dated February 9, 2025 - and it wasn't even published by AUTOMATIC1111 himself. The Issues section has discussion threads asking "Is this project dead?" [^1]. The community's center of gravity has long since shifted to ComfyUI, the node-based workflow engine iterating at a pace of one release per week [^2]. lllyasviel, the creator of ControlNet, started fresh with Forge, which can run FLUX on just 4GB of VRAM [^3].

The WebUI that once changed everything at 127.0.0.1:7860 now lies there quietly, like a veteran who has completed his mission.

This article is a complete memoir of it all. From the first line of code on that August afternoon in 2022, to the legacy it leaves behind today in 2026 - every version, every feature, every turning point, I will record as accurately as I can. Because some things, while you still remember them, should be written down.

Chapter 1 - August 22, 2022

To tell the story of A1111 properly, we have to go back to its cause - the birth of the Stable Diffusion model itself.

On August 22, 2022, Stability AI, in collaboration with the CompVis Lab at LMU Munich and Runway ML, publicly released the Stable Diffusion 1.4 model [^4]. It was a text-to-image generation model based on Latent Diffusion Model architecture, with approximately 890 million parameters, open-sourced under the CreativeML Open RAIL-M license. At the time, it was the most capable open-source image generation model, bar none. Its release sent shockwaves through the AI community akin to an earthquake.

But the model itself was just a set of weight files and an inference script. The official tool provided by Stability AI was a Python script that had to be run from the command line. For most people without a deep learning background - artists, designers, photography enthusiasts, curious netizens - "open a terminal, activate a conda environment, run python scripts/txt2img.py --prompt 'a cat'" was itself a barrier. And behind that barrier lay a series of obstacles that could make anyone give up at the first step: wrong CUDA version, PyTorch won't install, not enough VRAM, how to fill in the Hugging Face token...

On that very day, August 22, 2022, at 14:05:26 UTC, a GitHub user going by the name AUTOMATIC1111 created a public repository called stable-diffusion-webui [^5]. Ten minutes later, at 14:15:46, the first substantive code commit was pushed, with a commit message so terse it consisted of a single word: first [^6].

A web interface for Stable Diffusion, implemented using Gradio library.

• Repository description, August 22, 2022

This was no coincidence. A1111's birth and the release of Stable Diffusion 1.4 happened on the same day. Some online sources claim A1111 "appeared about a month after SD's release," which is inaccurate - the repository creation timestamp returned by the GitHub API is ironclad proof [^5][^6]. AUTOMATIC1111 responded the very day SD 1.4 was open-sourced, and this speed itself tells you something: he didn't act on impulse. He was prepared, waiting for the moment a model would be open-sourced.

And on that day, the democratization of AI image generation officially began.

Chapter 2 - A Spark from 4chan

Who is AUTOMATIC1111? This question still has no definitive answer. He has never revealed his real identity. His GitHub profile was created in 2022, and his commit records show the author name as AUTOMATIC with the email 16777216c@gmail.com [^6]. 16777216 is 2 to the 24th power, and the trailing c is widely believed to be a nod to 4chan - a connection made more explicit in the project's README.

Open the README file of the A1111 repository, and the Credits section reads:

Initial Gradio script - posted on 4chan by an Anonymous user. Thank you Anonymous user.

• stable-diffusion-webui README

This line reveals A1111's true starting point: AUTOMATIC1111 did not build this project from scratch. After Stable Diffusion 1.4 was released, an anonymous user on 4chan posted a simple Gradio-based script that let Stable Diffusion run in a browser. AUTOMATIC1111 took this script, used it as a starting point, rapidly expanded it into a fully-featured web application, and pushed it to GitHub [^4].

This origin story itself carries a distinctively internet-native character: an anonymous person, on an anonymous forum, posts anonymous code, and then another anonymous person picks it up and turns it into a product that changes the entire industry landscape. No company, no funding, no roadshow, no LinkedIn profile. Just a pseudonym on GitHub and a commit history.

There was also no affiliation with Stability AI. Stability AI was the upstream model provider; A1111 was an independent community project [^4]. This independence became critically important later - it meant A1111's fate was entirely tied to a single anonymous maintainer, which was both its greatest advantage (immune to commercial pressure) and its greatest risk (no organizational continuity guarantee).

On January 5, 2023, this risk manifested in dramatic fashion: AUTOMATIC1111's GitHub account was briefly banned for alleged Terms of Service violations, and the entire repository went offline, causing panic in the community [^4]. The ecosystem's core tool vanished along with its creator's account - the scenario every single-maintainer open-source project dreads most. The account was subsequently restored, and GitHub provided no detailed explanation, but the incident served as a wake-up call for everyone.

Ten days later, on January 15, 2023, AUTOMATIC1111 added the AGPL-3.0 open-source license to the project [^7]. For nearly five months prior, this project with tens of thousands of stars had no formal license file - the README even noted "Now with a license!" A project that reshaped the AI image generation landscape had been "running naked" legally for almost half a year. This sort of thing could probably only happen in this anonymous, decentralized community.

Chapter 3 - Five Months of Feral Growth

From August 22, 2022, to January 24, 2023 - nearly five months - A1111 never published a versioned Release. Users updated by running git pull to fetch the latest commits, and the changelog was the commit history [^4]. But it was precisely this "versionless" period that laid the foundation for nearly all of A1111's core features. The pace of development was so fast that "feral growth" is not an exaggeration.

On the day the repository was created, in addition to the basic txt2img function, AUTOMATIC1111 had already added GFPGAN face restoration and Prompt Matrix [^6]. The next day, August 23, prompt length validation and the --no-half command-line option were added, and img2img mode also gained Prompt Matrix support. On August 24, Textual Inversion support was merged, from a PR submitted by community contributor dogewanwan [^6].

Over the following two weeks, features grew at nearly one per day:

• 2022-09-03 - Inpainting: The commit message jokingly called it "poor man's inpainting" - a rudimentary but functional feature that let users select regions on existing images to regenerate.
• 2022-09-04 - ESRGAN Upscaling: A neural network image upscaler that allowed low-resolution generations to be enlarged. The same day also added Tiling and cross-attention layer optimization.
• 2022-09-04 - UI Config File: The introduction of ui-config.json allowed interface parameters to persist. The user experience evolved from "fill everything in again each time" to "remember my settings."
• 2022-09-30 - Embeddings Directory: Established a dedicated directory structure for Textual Inversion embedding files, paving the way for the later massive embedding ecosystem.
• 2022-10-08 - Hypernetwork Training: Inspired by the NovelAI model leak (October 2022), added Hypernetwork support and training templates.
• 2022-10-29 - Extensions Submodule: The embryonic form of the extension system appeared, planting the seed for the extension ecosystem that would later change the project's destiny.
• November 2022 - Localization and User Config: Added multilingual localization files and webui-user.bat, making it easy for non-English speakers and Windows users to launch.
• December 2022 - LoRA Support: Loading LoRA weights via the <lora:filename:multiplier> prompt syntax - this feature would become the cornerstone of the entire SD fine-tuned model ecosystem.

In two weeks, a script with only txt2img had transformed into a fully-featured image generation workbench: text-to-image, image-to-image, inpainting, face restoration, upscaling, textual inversion, hypernetwork training, prompt matrix, tiling, LoRA loading... this feature list would not be out of place today. And all of this was done without any version number, without a formal Release, without a changelog. AUTOMATIC1111 and the early community contributors, at a nearly frenetic pace, turned a 4chan script into a complete piece of software.

During this period, A1111 also established several conventions that would become de facto standards across the entire Stable Diffusion ecosystem. The three most far-reaching were:

Three Legacy Conventions

• Negative Prompt: In addition to the positive prompt, a separate negative prompt input field was added, telling the model "what not to generate." This design seems obvious in hindsight, but before A1111, most SD inference scripts did not treat it as an independent, first-class input field.
• Prompt Weighting Syntax: Syntax like (word:1.2) let users assign different weights to specific words in the prompt, precisely controlling generation direction. This syntax was later adopted as standard by virtually all model-sharing platforms, including Civitai and Hugging Face.
• PNG Metadata Embedding: Generated PNG images automatically embedded complete generation parameters (model, prompt, negative prompt, sampler, steps, CFG scale, seed, etc.). This meant any image generated by A1111 could be reproduced by "dragging it back into the interface." This convention was also widely adopted across the ecosystem.

None of these three conventions were technically complex. But together they constituted a "user experience standard" - when everyone was using A1111, A1111's way of doing things became the standard. Model description pages on Civitai were written according to A1111's usage, tutorials used A1111's interface screenshots, and prompts followed A1111's syntax format. This ecosystem lock-in effect is one of A1111's most profound legacies [^4].

Chapter 4 - Year One of Version Numbers

On January 24, 2023, A1111 finally released its first tagged version: v1.0.0-pre [^4]. The -pre suffix betrayed a modest humility - even with tens of thousands of stars and a complete feature set, AUTOMATIC1111 still considered this merely a "pre-release."

The significance of v1.0.0-pre lay not in features - by then, txt2img, img2img, inpainting, upscalers, textual inversion, hypernetworks, LoRA, xformers, --medvram / --lowvram, and other core features had long existed - but in formalizing the extension system. From this version on, A1111 provided a standard extension installation mechanism: users could install third-party extensions by entering a GitHub URL in the Extensions tab, or use built-in extensions from the extensions-builtin/ directory [^8].

Additionally, this version provided a sd.webui.zip binary distribution package for users who couldn't even install Python and Git - just extract and double-click run.bat to run (Windows 10 + NVIDIA GPU only) [^9]. At a time when AI image generation tutorials were still stuck on "how to configure a conda environment," this one-click launcher dramatically lowered the barrier to entry.

From v1.0.0-pre onward, A1111 entered a rhythmic release cycle. Over the next roughly year and a half, from v1.1.0 to v1.10.0, eleven major versions were released, each bringing substantial feature updates and architectural improvements. This period was A1111's golden age.

Chapter 5 - A Chronicle of Features

Below is the complete version chronicle of A1111 from v1.1.0 to v1.10.1. All dates are from the official GitHub Release pages [^9].

v1.1.0 - May 1, 2023

The first formal