Stable Diffusion Tutorial 2026: Install and Run It Locally

Stable Diffusion is an open-weight AI image generator that creates images from text prompts. The current release, Stable Diffusion 3.5, comes in Large (8.1 billion parameters), Large Turbo and Medium (2.5 billion parameters, MMDiT-X architecture) variants, while SDXL and SD1.5 remain popular for their extensive model ecosystems and lower hardware requirements.

Key facts

Updated 2026-09-24: added Key facts, question headings with direct answers, and an FAQ built from the questions Google shows for this topic.

What is the latest Stable Diffusion version in 2026?

Stable Diffusion 3.5, in Large, Large Turbo, and Medium variants, is the latest official Stable Diffusion release as of September 2026. Stability AI's 2026 launches focused elsewhere: Stable Audio 3.0 shipped in May 2026 and the Brand Studio platform in April 2026 ↗. No SD 4 has been announced.

Updated September 2026. The practical model landscape:

• SD 3.5 Large (8.1B): highest-quality official model, MMDiT architecture, strong text rendering

• SD 3.5 Large Turbo: distilled version of Large that generates in 4 steps for fast iteration

• SD 3.5 Medium (2.5B): designed to run on consumer GPUs

• SDXL: the largest fine-tune and LoRA ecosystem; the default for community workflows

• FLUX.2: popular for photorealism, from Black Forest Labs, whose founding team includes the researchers behind Latent Diffusion and Stable Diffusion

New checkpoints, LoRAs, and ControlNets are published regularly for SDXL and Flux, while Stability's own roadmap has shifted toward enterprise tools. Pick SD 3.5 for official support and prompt adherence, SDXL for ecosystem depth.

What's new in Stable Diffusion 3.5?

Stable Diffusion 3.5 is a family of MMDiT models with better text rendering, prompt adherence and customizability than earlier SD 3 releases. Large has 8.1 billion parameters. Medium uses the improved MMDiT-X architecture at 2.5 billion parameters and needs 9.9GB of VRAM, excluding text encoders, according to Stability AI's announcement.

Stable Diffusion 3.5 includes these key improvements, according to Stability AI's announcement ↗:

• MMDiT-X Architecture (Medium): Improved architecture and training methods for multi-resolution generation between 0.25 and 2 megapixels

• SD 3.5 Medium: Needs 9.9GB of VRAM (excluding text encoders) for full performance, per Stability AI

• Text Rendering: Improved typography, i.e. more readable text within images

• Customizability: Query-Key normalization in the transformer blocks makes fine-tuning more stable

• Natural Language Processing: Better handling of complex prompts, including spatial relationships like "left," "right," and "behind"

Which Stable Diffusion version should I use?

Use SD 3.5 for prompt adherence and text rendering, SDXL for the largest ecosystem of community checkpoints and LoRAs, and SD 1.5 for low-VRAM systems. SD 3.5 Medium needs 9.9GB of VRAM excluding text encoders, SDXL is comfortable on 8GB, and SD 1.5 runs on about 4GB with low-VRAM options.

The Stable Diffusion ecosystem includes three main versions:

SD 3.5 (Latest)

• Resolution: about 1 megapixel (1024x1024) for Large; 0.25 to 2 megapixels for Medium

• Parameters: 8.1 billion (Large), 2.5 billion (Medium)

• VRAM requirement: 9.9GB for Medium excluding text encoders; Large needs more

• Best for: Photorealistic images, text rendering, complex prompts

SDXL (Most Popular)

• Resolution: 1024x1024 pixels

• Parameters: 3.5 billion (base model)

• VRAM requirement: 8GB is comfortable; less works with memory-saving options

• Best for: Artistic styles, custom models, LoRA training

SD 1.5 (Lightweight)

• Resolution: 512x512 pixels

• Parameters: 860 million (UNet)

• VRAM requirement: about 4GB with Automatic1111's low-VRAM options

• Best for: Fast generation, older hardware, largest model selection

How do I install Stable Diffusion?

ComfyUI, Automatic1111 and Forge WebUI are the three main interfaces for running Stable Diffusion locally. Each is a git clone plus a Python install: ComfyUI's portable build ships with Python 3.13, while Automatic1111 asks Windows users for Python 3.10.6. You also need a GPU with enough VRAM for the model you choose.

ComfyUI Installation (Recommended)

git clone https://github.com/Comfy-Org/ComfyUI ↗
cd ComfyUI
pip install -r requirements.txt
python main.py
ComfyUI provides node-based workflow creation and supports SD 1.5, SDXL and SD 3.5, plus Flux and other newer models. Install PyTorch for your GPU first; the README recommends Python 3.13.

Automatic1111 Installation

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui ↗
cd stable-diffusion-webui
./webui.sh # Linux/Mac
webui-user.bat # Windows
Automatic1111 offers a traditional web interface with extensive extension support. Its README asks Windows users to install Python 3.10.6, and the last tagged release is v1.10.1 (February 2025).

Forge WebUI Installation

git clone https://github.com/lllyasviel/stable-diffusion-webui-forge ↗
cd stable-diffusion-webui-forge
webui-user.bat # Windows (./webui.sh on Linux)
Forge WebUI is built on Automatic1111 with better memory management and faster inference, and adds Flux support including NF4 and GGUF quantized models. Windows users can also use its one-click package.

What hardware do I need for Stable Diffusion?

Needs depend on the model: SD 1.5 can run on about 4GB of VRAM, SDXL is comfortable on 8GB, and Stability AI lists 9.9GB of VRAM (excluding text encoders) for SD 3.5 Medium. More VRAM means faster 1024x1024 generation and room for larger models.

The tiers below are rough guidance rather than official requirements.

Minimum Requirements:

• GPU: RTX 3060 (12GB version) or RTX 4060

• RAM: 16GB system memory

• Storage: 50GB free space for models

Recommended Setup:

• GPU: RTX 4070 Ti (12GB VRAM) or RTX 3080

• RAM: 32GB system memory

• Storage: SSD with room for several large checkpoints

Optimal Configuration:

• GPU: RTX 4090 (24GB VRAM) or RTX 3090

• RAM: 64GB system memory

• Storage: 1TB NVMe SSD

Mac Users:

• Apple Silicon (M-series) Macs run ComfyUI through PyTorch's MPS backend; 16GB+ unified memory is advisable

• Expect slower generation than on a comparable NVIDIA GPU

How do I write effective prompts for SD 3.5?

SD 3.5 processes natural language prompts better than keyword lists, understanding spatial relationships and generating readable text with descriptive sentences rather than comma-separated tags.

Prompt Structure for SD 3.5

SD 3.5 responds to descriptive sentences:

A professional photo of a woman in a red dress standing in a sunlit garden, shallow depth of field, golden hour lighting, shot on Sony A7III

Key Prompt Improvements

Natural Language Processing:

• Write prompts as complete descriptions

• Use spatial terms: "woman standing behind the tree"

• Include camera settings: "shot on Canon 5D, 85mm lens"

Text Rendering:

• Specify text content: "sign reading 'OPEN'"

• Include text style: "bold red letters on white background"

• Position text: "text centered at bottom of image"

Style Consistency:

• Describe artistic style: "painted in impressionist style"

• Specify medium: "oil painting on canvas"

• Include lighting: "dramatic chiaroscuro lighting"

Why do artists still use SDXL and SD 1.5?

SDXL has a very large library of fine-tuned checkpoints and LoRAs on sites like Civitai, while SD 1.5 runs on low-VRAM systems and generates quickly at 512x512.

SDXL Advantages

Model Ecosystem:

• A large number of fine-tuned checkpoints on Civitai

• A large LoRA (Low-Rank Adaptation) library

• Extensive style variety from photorealistic to anime

• Active community publishing new models regularly

Training Flexibility:

• LoRA training runs on consumer GPUs with tools like Kohya_ss

• DreamBooth fine-tuning supports custom subjects

• Textual Inversion creates new concepts

• ControlNet provides precise pose and composition control

SD 1.5 Benefits

Hardware Efficiency:

• Runs on 4GB VRAM GPUs with Automatic1111's low-VRAM options

• Generates 512x512 images faster than SDXL on the same GPU

• Uses less memory than SDXL

• Handles larger batches on the same hardware

Model Variety:

• A huge back catalog of community models

• Specialized models for anime, photography, art styles

• Many techniques first appeared for SD 1.5

• Most comprehensive tutorial coverage

What advanced techniques work with SD 3.5?

SD 3.5 Large has official ControlNets, regional prompting is available through interface tools, and multi-stage workflows can combine different SD versions.

ControlNet Integration

Stability AI released official Blur, Canny and Depth ControlNets for SD 3.5 Large ↗. Other control types are more mature on SDXL and SD 1.5:

Pose Control:

• OpenPose detects human body keypoints

• DWPose improves hand and face detection

• Animal pose detection for pets and wildlife

Depth Control:

• Depth maps (the official SD 3.5 Large depth ControlNet uses DepthFM) guide 3D scene layout

• Depth maps control foreground/background separation

• Perspective guidance for architectural images

Edge Control:

• Canny edge detection preserves line art

• Scribble control allows rough sketch input

• Normal maps create detailed surface textures

Regional Prompting

SD 3.5 has no native regional prompt syntax. Regional prompting comes from interface tools, such as area and mask conditioning nodes in ComfyUI or regional-prompt extensions for WebUI. Conceptually, each zone gets its own prompt:

[left side: red roses in bloom]
[right side: blue violets in grass]
[background: golden sunset sky]
Regional prompting divides images into zones with specific prompts for each area.

Multi-Stage Workflows

Advanced workflows can combine multiple SD versions:

1. Base Generation: SD 3.5 for prompt adherence and composition

2. Style Refinement: SDXL for artistic enhancement and detail

3. Specialized Details: SD 1.5 models for specific elements

4. Upscaling: ESRGAN or Real-ESRGAN for resolution increase

Which models should I download in 2026?

SD3.5 Medium is the lightest official SD 3.5 model, JuggernautXL is a popular photorealistic SDXL checkpoint, and DreamShaper XL is a popular choice for artistic and fantasy content.

SD 3.5 Models

SD3.5-Medium:

• File size: 5.1GB (sd3.5_medium.safetensors on Hugging Face)

• Parameters: 2.5 billion

• VRAM usage: 9.9GB excluding text encoders, per Stability AI

• Best for: General purpose image generation

SD3.5-Large:

• File size: 16.5GB (sd3.5_large.safetensors)

• Parameters: 8.1 billion

• VRAM usage: higher than Medium; quantized versions reduce it

• Best for: Maximum quality output

SD3.5-Large-Turbo:

• File size: 16.5GB (sd3.5_large_turbo.safetensors)

• Generation steps: 4 (Adversarial Diffusion Distillation)

• Speed: considerably faster than SD 3.5 Large

• Best for: Rapid iteration and previews

Popular SDXL Models

JuggernautXL:

• Specialization: Photorealistic humans and portraits

• Source: Civitai

• File size: standard full SDXL checkpoint

• Best for: Headshots and portraits

DreamShaper XL:

• Specialization: Artistic and fantasy imagery

• Source: Civitai

• File size: standard full SDXL checkpoint

• Best for: Creative and stylized artwork

RealVisXL:

• Specialization: General-purpose photorealism

• Source: Civitai

• File size: standard full SDXL checkpoint

• Best for: Realistic scenes and products

What tools enhance Stable Diffusion workflows?

ComfyUI Manager simplifies custom node and model installation, Civitai hosts a very large library of community models, Kohya_ss GUI enables LoRA training, and 4x-UltraSharp is a widely used upscaling model.

Essential Management Tools

ComfyUI Manager:

• Model downloads from within ComfyUI

• Missing node detection for downloaded workflows

• Update management for installed custom nodes

• Custom node management for extensions

Civitai Platform:

• Large library of checkpoints and LoRAs

• User ratings and reviews for quality assessment

• Model version tracking and updates

• Downloads usable in local interfaces

Training and Customization

Kohya_ss GUI:

• LoRA training interface for custom subjects

• DreamBooth fine-tuning capabilities

• Settings for different GPU types

• Training time depends on dataset size, resolution and GPU

Dataset Preparation:

• A few dozen varied images is a common starting point for LoRA training

• Train near the model's native resolution (512x512 for SD 1.5, 1024x1024 for SDXL)

• Automatic tagging with BLIP or WD14 taggers

• Data augmentation for improved training results

Upscaling Solutions

4x-UltraSharp:

• 4x upscaling model (e.g., 512x512 to 2048x2048)

• Preserves fine details and textures

• Processing time depends on image size and GPU

• Best for: General purpose upscaling

Real-ESRGAN:

• Specialized for photographic content

• Multiple model variants for different content types

• Batch processing support

• Best for: Realistic image enhancement

How do I solve common SD 3.5 problems?

SD 3.5 needs more VRAM than SDXL and responds better to descriptive prompts than keyword lists, while performance optimization includes attention optimizations and tiled generation for large images.

SD 3.5 Specific Issues

Smaller Fine-tuning Ecosystem:

• LoRA training for SD 3 models is supported in tools such as Hugging Face diffusers ↗

• The library of community SD 3.5 LoRAs and checkpoints is much smaller than SDXL's

• Current workaround: Use SDXL for custom training when you need a specific community style

Higher VRAM Usage:

• Enable tiled generation for images larger than 1024x1024

• Use model offloading to system RAM when VRAM insufficient

• Reduce batch size from 4 to 1-2 images per generation

Prompt Style Differences:

• Replace keyword lists with descriptive sentences

• Use natural language instead of booru tags

• Include camera settings and lighting descriptions

Performance Optimization

Memory Management:

• Enable xformers or PyTorch SDP attention for faster, leaner attention

• Use channels-last memory format for efficiency

• Use fp16 or quantized weights for VRAM savings

Generation Optimization:

• Batch multiple images when VRAM allows

• Use pruned fp16 checkpoints to reduce file sizes

• Enable CPU offloading for systems with limited VRAM

Quality Settings:

• Start around 28 steps for SD 3.5 Large and 40 for Medium (Stability's example settings)

• Use a low CFG/guidance scale for SD 3.5 (3.5 for Large, 4.5 for Medium in the official examples); 7-9 is typical for SD 1.5/SDXL

• DPM++ 2M Karras is a common quality/speed choice for SD 1.5 and SDXL

What's coming next for Stable Diffusion?

As of September 2026, Stability AI has not announced an SD 4.0 or a newer open image model. Its 2026 announcements covered Stable Audio 3.0 and the Brand Studio platform.

Community Developments

Licensing Concerns:

• SD 3.5 uses the Stability AI Community License: free for non-commercial use and for commercial use by organizations under $1M annual revenue

• SD 1.5 and SDXL use CreativeML OpenRAIL licenses, which allow commercial use with use-based restrictions

• Alternative model development (for example FLUX.2 and other open-weight models) is increasing

Technical Evolution:

• Distilled few-step models (like SD 3.5 Large Turbo) for faster generation

• Mobile optimization for smartphone deployment

• Edge computing integration for local processing

• API standardization across different implementations

FAQ

How do you run Stable Diffusion?

Install one of the three local interfaces, then download a model checkpoint. ComfyUI is a git clone, a pip install and python main.py; Automatic1111 and Forge start from webui.sh or webui-user.bat. Put the checkpoint in the models folder, open the local web address the interface prints, write a prompt and generate.

What do you need to use Stable Diffusion?

A computer with Python and a GPU sized for the model: about 4GB of VRAM for SD 1.5, 8GB for SDXL, and 9.9GB excluding text encoders for SD 3.5 Medium. Apple Silicon Macs work through PyTorch's MPS backend but run slower. Budget roughly 50GB of disk for checkpoints, more if you collect SDXL models.

Is Stable Diffusion still worth it?

Yes for local, no-fee generation and custom training, with caveats. Stability AI has not announced an SD 4, and its 2026 releases were Stable Audio 3.0 and Brand Studio. SDXL still has the largest checkpoint and LoRA ecosystem, and FLUX.2 is the popular open-weight alternative for photorealism, so the ecosystem keeps moving even without a new SD model.

Can we use Stable Diffusion for free?

Yes. The models are free to download and run locally, so the only costs are hardware and electricity. SD 1.5 and SDXL use CreativeML OpenRAIL licenses with use-based restrictions. SD 3.5 is free for non-commercial use, and for commercial use by organizations under US $1,000,000 in annual revenue, under the Stability AI Community License.

Can I use Stable Diffusion commercially?

Yes, with conditions. SD 1.5 and SDXL use CreativeML OpenRAIL licenses that permit commercial use with use-based restrictions. SD 3.5 is free for commercial use under the Stability AI Community License ↗ if your organization's annual revenue is under $1M; larger companies need an enterprise license.

How much does it cost to run Stable Diffusion?

The models are free to download and run locally. Your costs are the GPU hardware (or cloud GPU rental) and electricity.

Can I train custom models on my own images?

Yes. LoRA training with tools like Kohya_ss or diffusers is the usual starting point and runs on consumer GPUs. Full DreamBooth fine-tuning needs more images, VRAM and time.

Which is better: Stable Diffusion or DALL-E 3?

DALL-E 3 was removed from the OpenAI API on May 12, 2026, per OpenAI's deprecations page ↗; OpenAI's GPT Image models replaced it. Those offer easier prompting and ChatGPT integration but are paid per image via the API. Stable Diffusion provides local generation with no per-image fee, custom model training, and full creative control after the hardware investment.

Can Stable Diffusion generate NSFW content?

Local open-weight models have no mandatory filter beyond what your interface adds, but the model licenses and Stability AI's Acceptable Use Policy restrict certain content. Commercial platforms typically restrict NSFW generation regardless of the underlying model.

How do I improve image quality?

Use higher resolution models (SDXL/SD 3.5), use the step and guidance settings recommended for your model, apply upscaling with Real-ESRGAN, use negative prompts to exclude unwanted elements, and experiment with different samplers like DPM++ 2M Karras.

Sources

• Introducing Stable Diffusion 3.5, Stability AI ↗

• Stability AI news (Stable Audio 3.0, Brand Studio), Stability AI ↗

• Stability AI Community License, Stability AI ↗

• Stable Diffusion 3.5 Large model card, Hugging Face ↗

• Stable Diffusion 3.5 ControlNets, Hugging Face ↗

• SDXL 1.0 announcement, Stability AI ↗

• Stable Diffusion WebUI (Automatic1111), GitHub ↗

• ComfyUI repository, Comfy-Org on GitHub ↗