Stable Diffusion Review 2026

What is Stable Diffusion?

Most tutorials that answer "Stable Diffusion: how to install locally on PC" were written around 2023, when the AUTOMATIC1111 web UI and SD 1.5 were what everyone ran. A lot of that advice has aged badly. Stability AI's current open-weight family is Stable Diffusion 3.5 (Large, Large Turbo and Medium), AUTOMATIC1111 has not shipped a feature release since its 1.10 series, and ComfyUI now offers a desktop installer for Windows and macOS. Below: what to download, the hardware it needs, the licence, and when paying for Midjourney or the API is smarter.

Think of it as a set of downloadable image models rather than one app. Stability AI, the London company founded in 2019 that released the first version in August 2022, publishes the base weights; the interfaces, fine-tunes and add-ons come from a large community. That split explains both the appeal (no subscription, full privacy) and the friction (you are your own IT department).

Which Stable Diffusion model to download in 2026

Stability's image page lists five models, with the 3.5 releases newest:

• SD 3.5 Large: 8.1 billion parameters, designed for 1 megapixel output, pitched as the family's best for quality and prompt adherence.
• SD 3.5 Large Turbo: a distilled version that finishes an image in four steps.
• SD 3.5 Medium: 2.5 billion parameters per Stability's launch post (the Hugging Face card shows 2B), on an MMDiT-X design, for images between 0.25 and 2 megapixels.
• SDXL 1.0 and SDXL Turbo: older, but with the deepest pool of community checkpoints and LoRA files.

With a mid-range graphics card, begin with SD 3.5 Medium or SDXL and move to Large later.

Stable Diffusion system requirements: GPU, VRAM and RAM for each model

Official VRAM numbers are thin. Stability's launch post says SD 3.5 Medium needs 9.9 GB of VRAM, and that figure excludes the text encoders. The Hugging Face cards for Medium and Large give no minimum; they show 4-bit NF4 loading with CPU offloading for smaller cards. For Large, our own estimate: 8.1 billion parameters in BF16 come to roughly 16 GB for the weights alone, before the three text encoders (two CLIP models and T5-XXL).

NVIDIA remains the path of least resistance. ComfyUI's portable Windows build targets RTX 20-series and newer, with a separate build for older 10-series cards. AMD works through ROCm on Linux and on Windows 11, Intel Arc through PyTorch's XPU backend, and Apple Silicon Macs are supported. ComfyUI's README claims even the biggest open models can run on 4 GB of VRAM plus 8 GB of RAM, a vendor claim without published benchmarks.

Stable Diffusion: how to install locally on PC with ComfyUI Desktop

• Install the ComfyUI Desktop app for Windows or macOS from the link on its GitHub page. Linux users follow the README's manual install.
• Create a free Hugging Face account. The SD 3.5 repositories are gated, so you accept the Stability Community License and share contact details before downloads open up.
• Download the model file and the text encoder files listed on its card. The Large repository includes an FP8 build of T5-XXL to save memory.
• Put the checkpoint in ComfyUI's models/checkpoints folder and the encoders in the text encoder folder, then open the example workflow linked from the model card.
• Run a test prompt at 1024x1024. The Large card's examples use 28 steps and a guidance scale of 3.5 to 4.5.
• Add custom nodes and LoRA files one at a time.

The older route still suits SDXL and SD 1.5. Install Python 3.10.6 (the AUTOMATIC1111 README warns newer Python versions do not support its torch build), install Git, clone the repository and double-click webui-user.bat.

Stable Diffusion AUTOMATIC1111 vs ComfyUI: which one is still moving

AUTOMATIC1111's web UI is what most tutorials still show: a tabbed form with txt2img, img2img, inpainting, X/Y/Z plots and a built-in API. It has around 165k GitHub stars and an AGPL-3.0 licence. The problem is pace. Its latest tagged release is 1.10.1, the 1.10.0 notes added basic SD 3 support (Euler sampler recommended, other timestep samplers unsupported, T5 off by default), and 1.10.1 was a single bug fix. SD 3.5 is not mentioned in the release history. Roughly 2,400 issues sit open.

ComfyUI, now under the Comfy-Org organisation, takes the opposite approach. You wire nodes into a graph: every step is visible, workflows save as shareable files, and new models tend to work there first. Its README lists SD 1.5, SDXL, SD 3.5, Flux.1, Flux.2, Qwen Image and Z-Image for images, plus Wan and LTX-Video for video. The README promises a new stable version roughly every two weeks, the licence is GPL-3.0, and the repo has about 140k stars.

Our read: start with ComfyUI, and keep AUTOMATIC1111 only for an existing SDXL setup.

LoRA, ControlNet and inpainting: why people put up with the setup

The base model is only the starting point. A LoRA is a small add-on file, trained on a few dozen images of a product, face or art style, that you load on top of a checkpoint to steer results without retraining the whole model. ControlNet takes a guide image (a pose skeleton, a depth map, edge lines) and makes the composition follow it, useful when a product must sit in the same spot across many variants. Inpainting repaints a masked area, such as a bad hand or a background.

All three are tied to a model family. A LoRA or ControlNet trained for SDXL will not work properly on SD 3.5, so check the base model tag on Civitai or Hugging Face before you download. Stability's licence treats LoRAs and fine-tunes as derivative works rather than new foundation models, so you may train and even sell them under the same revenue rule described below.

The $1M line in Stability's Community License

The SD 3.5 models ship under the Stability AI Community License. According to stability.ai/license, use is free for research, non-commercial projects and commercial work by individuals or organisations with less than $1M in total annual revenue, counted from every source, AI or otherwise. Cross that line and you need an Enterprise License, priced on request. Between you and Stability, you own the images you generate, subject to local law and the Acceptable Use Policy.

SDXL 1.0 base follows different rules: its Hugging Face card lists the CreativeML Open RAIL++-M licence, which has no revenue cap but does carry use restrictions. For a freelancer or a startup under $1M, both are effectively free for client work. Copyright protection for purely AI-made images varies by country.

Paying per image through Stability's API credits

No usable GPU? Stability sells the same models through its API on credits. Its pricing page sets one credit at $0.01 (so $10 buys 1,000 credits) and offers 25 free credits to start, while warning that prices can change.

• SD 3.5 Large: 6.5 credits ($0.065) per image
• SD 3.5 Large Turbo: 4 credits ($0.04)
• SD 3.5 Medium: 3.5 credits ($0.035)
• SD 3.5 Flash: 2.5 credits; Stable Image Core: 3; Stable Image Ultra: 8
• Inpaint, Erase Object and Remove Background: 5 credits each

No rupee price is listed. At $0.035 an image, 1,000 Medium images cost $35, which is the figure to set against a Midjourney plan or a GPU upgrade. Stability's privacy policy says inputs and outputs from its hosted services may be used to train models, with an opt-out in account settings. Generating locally sends nothing anywhere.

Stable Diffusion free online: what is left after DreamStudio

DreamStudio, Stability's original browser app, is gone as a consumer product: dreamstudio.ai now redirects to brandstudio.com. Stability's image page points browser users to Stable Assistant instead, so check its plans before assuming it is free. The genuinely free routes are a Hugging Face Space linked from the SD 3.5 Large card (queues vary), the 25 starter API credits, and third-party hosts. Treat random free generator sites with suspicion. Comfy Cloud runs ComfyUI without a GPU, but it is paid.

Stable Diffusion vs Midjourney for a small studio

Midjourney is the easier tool. Its plans cost $10 (Basic), $30 (Standard), $60 (Pro) and $120 (Mega) a month, 20% less when billed yearly, and its plans page mentions no free trial. Stealth mode, which keeps your images out of the public gallery, begins at Pro, and companies with over $1M in yearly gross revenue must be on Pro or Mega.

Local generation wins on control and on cost at volume. You can train a LoRA on a client's product, lock a layout with ControlNet and queue hundreds of variations overnight without paying per image. Midjourney wins on time to a good-looking image. An agency can sensibly use Midjourney for mood boards and a local pipeline for production work. Leonardo AI and Adobe Firefly are the other realistic alternatives, and Flux models from Black Forest Labs, which also run inside ComfyUI, are now the main open-weight rival.

Is running it yourself safe, and where it bites

• Custom nodes and extensions are Python code from strangers with full access to your PC. Stick to popular, maintained ones.
• Prefer .safetensors files over older .ckpt checkpoints, which can carry executable code.
• No support desk for local use; help comes from GitHub issues and forums.
• A GPU with 12 to 16 GB of VRAM is a real expense for an Indian freelancer, and laptops with 6 or 8 GB cards will struggle with 3.5 Large.
• Stability warns that vague prompts give less predictable SD 3.5 results; many SDXL fine-tunes still expect tags plus a negative prompt.

A first weekend with SD 3.5 Medium

On day one, install ComfyUI Desktop, download SD 3.5 Medium with its encoders and generate 20 images you actually need, such as a product shot or a blog header, noting the seconds per image. On day two, try inpainting on your best result, then load one LoRA from Civitai built for the same base model. Usable images in under a minute each? Local generation earns its place. If not, run the same prompts through the API at 3.5 credits apiece, or pay for one month of Midjourney Basic, and compare the bills.