What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
ThatPainter is reader-supported. When you buy through links on our site, we may earn an affiliate commission. Learn More
To use Stable Diffusion 3 Medium in SwarmUI, install SwarmUI with ComfyUI as the backend, add the SD3 Medium model, let its required text encoders download, and begin with controlled settings rather than chasing a magic preset. Start with the model card’s example of 28 steps and guidance scale 7, keep Sigma Shift at 3, use CLIP-only text encoding when resources are limited, and enable Refiner Do Tiling for SD3 upscaling.
SwarmUI is not a new image model. It is the interface and workflow manager that makes model selection, prompting, parameter changes, grids, history, and backend orchestration easier to manage. This distinction is the key to getting reliable results instead of treating the UI as a quality switch.
What SwarmUI and Stable Diffusion 3 each do
The first distinction matters: Stable Diffusion 3 is the image-generation model family; SwarmUI is the interface and workflow manager. In a typical local setup, SwarmUI provides the web interface for prompts, model selection, parameters, image history, grids, and generation controls, while a backend such as ComfyUI performs the underlying workflow. The official SwarmUI project describes it as a modular Stable Diffusion web UI focused on accessible power tools and extensibility.
#1 Best Overall
- Axial-tech fans now feature a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- 2.5-slot design allows for greater build compatibility while maintaining cooling performance
- 0dB technology lets you enjoy light gaming in relative silence
- Dual BIOS switch lets you toggle between Quiet and Performance BIOS profiles
- Dual ball fan bearings last up to twice as long as sleeve bearing designs
For this guide, the practical target is Stable Diffusion 3 Medium, the locally downloadable SD3 model most directly documented for this workflow. SD3 itself uses a diffusion-transformer architecture with flow matching, but installing SwarmUI does not change the model or automatically improve its image quality. Results still depend on the model files, prompt, text-encoder mode, sampler, resolution, steps, guidance, seed, hardware, and refinement workflow.
The short version: install SwarmUI, select ComfyUI as its local backend, add Stable Diffusion 3 Medium through the normal model workflow, allow the required text encoders to download, and begin with the SD3 model card’s illustrative baseline of 28 steps and guidance scale 7. Keep Sigma Shift at its documented default of 3, use CLIP-only text encoding on constrained hardware, and enable Refiner Do Tiling before upscaling. These are sensible starting points, not universal best settings.
Before you install: check the machine
SD3 Medium is demanding because the pipeline may use multiple large components, including text encoders. The official Diffusers documentation warns that SD3 can be challenging on GPUs with less than 24 GB of VRAM even with FP16. That is a practical warning, not a minimum hardware law: lower-VRAM systems may work with memory-saving settings, offloading, quantization, reduced text-encoder use, lower resolutions, or a slower workflow.
| Machine category | What to expect |
|---|---|
| NVIDIA GPU with 24 GB or more of VRAM | The most defensible local target category for experimenting with SD3 Medium at useful settings. It is not a claim that one particular card is mandatory. |
| NVIDIA GPU below 24 GB | Possible in some configurations, but memory pressure, offloading, reduced settings, and long generation times become more likely. Do not assume 12 GB or 16 GB will be comfortable for every workflow. |
| AMD GPU | Compatibility depends on the operating system, drivers, PyTorch, and ROCm support. SwarmUI’s troubleshooting documentation notes that AMD ROCm on Windows has historically been limited and may perform better under Linux; treat this as a current compatibility consideration, not a permanent rule. |
| Apple-silicon Mac | SwarmUI’s project documentation supports Macs with Apple M-series processors. Performance and memory behavior will not be identical to an NVIDIA CUDA system. |
| No suitable local GPU | Consider hosted inference or a hosted GPU instead of treating SwarmUI as a way to bypass the model’s memory requirements. This guide focuses on local installation. |
Do not confuse GPU VRAM with ordinary system RAM. VRAM is the main constraint for loading and running the diffusion pipeline on the graphics card. System memory can help with multitasking or offload-heavy configurations, but a desktop RAM upgrade does not substitute for insufficient VRAM.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Storage is also worth planning. Model weights, text encoders, caches, generated images, grids, and upscaled versions can consume substantial space. A high-capacity NVMe SSD is an optional convenience for model files and outputs, not a SwarmUI requirement or a guarantee of faster generation once the workload is GPU-bound.
Disclosure: Product references in this guide are optional equipment ideas, not mandatory purchases. Product fit, price, availability, and any applicable product-link relationship should be verified before buying.
Install SwarmUI on Windows, Linux, or Apple silicon
Use the installation instructions in the official SwarmUI repository rather than an unofficial repackaged installer. The project supports Windows, Linux, and Apple-silicon Macs, but the prerequisites and performance characteristics differ by platform.
Windows
- Download SwarmUI using the official project’s Windows installation or launch instructions.
- Prefer a normal internal NTFS drive for the installation. SwarmUI’s troubleshooting documentation associates Git ownership errors with installations on unsuitable drives and specifically recommends a standard internal NTFS drive on Windows.
- Run the official installer or launch process and wait for the on-page setup interface.
- When the installer asks for a backend, select ComfyUI for the usual local beginner workflow.
- Choose local or private access for a single-computer installation. Do not expose the server to the internet simply to access it from another device.
Linux
Prepare Git, a supported .NET installation, and Python in the range documented by the project. The current SwarmUI documentation recommends Python 3.10 through 3.12 and warns against Python 3.13 for this setup. Check the current README immediately before installing because supported dependency versions can change.
After the prerequisites are available, follow the repository’s Linux launch instructions. Use the setup choices to select ComfyUI as the backend and keep the initial access mode private. Avoid mixing system-wide Python packages with the environment created for SwarmUI.
Apple-silicon Mac
The project’s Mac instructions target Apple M-series processors. Follow the current repository instructions for the supported launch process and expect different speed, memory behavior, and backend support from an NVIDIA CUDA workstation. Apple-silicon support means the platform is supported; it does not promise parity with a 24 GB NVIDIA GPU.
Make the initial installer choices deliberately
The SwarmUI installer asks you to select a theme, access mode, backend, and starter model. For a first local installation:
- Choose any theme you can read comfortably; it has no bearing on image quality.
- Choose a private or local access mode unless you have a specific, secured multi-device plan.
- Choose ComfyUI as the backend for the normal documented local arrangement.
- Select a starter model if the installer offers one, or add SD3 Medium after the interface is running.
Once setup finishes, open the local SwarmUI address shown by the launcher. The exact appearance can change between releases, so use the current interface labels rather than relying on screenshots from an older build.
Recommended Free Tools
Secure the interface before sharing it
A local-only interface and a public web service are different projects. If you intend to access SwarmUI from outside the machine or host it publicly, follow the official advanced-usage guidance: configure authentication and HTTPS.
This is not merely a general best practice. The documentation warns that some API routes can forward raw ComfyUI access, which creates an abuse risk on an unauthenticated public instance. Do not port-forward a default local server and assume an obscure URL is protection.
Load Stable Diffusion 3 Medium
SwarmUI’s model-support documentation lists Stable Diffusion 3 Medium as supported through the normal model workflow. Obtain the model from an authorized source and read the current SD3 Medium model card before downloading or publishing work made with it.
- Start SwarmUI with the ComfyUI backend. Confirm that the backend finishes loading before troubleshooting the model itself.
- Add or select the SD3 Medium model. If you are downloading it through the supported workflow, allow the process to finish rather than interrupting it partway through.
- Allow the first-run text-encoder download. SwarmUI can download the required SD3 text encoders on the first SD3 generation. Ensure you have network access and enough free storage.
- Return to the main generation page. The model should be available in the model selector once its files and folder mapping are recognized.
If the model is not visible
First verify that the model was downloaded completely. Then check the model root and folder mapping. SwarmUI documents configuration for existing Auto1111 and ComfyUI model directories in its basic-usage documentation. A model stored in a directory that the active backend does not scan will not appear in the selector.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsDo not immediately install a collection of custom nodes or manually copy packages to compensate. Confirm the active backend, model directory, file placement, and startup messages first.
Rank #2
- 【Ultimate Triple Display Connectivity】: Features a versatile output array including HDMI, DisplayPort (DP), and DVI. Whether you're connecting a high-refresh-rate gaming monitor via DP or a standard office screen via HDMI, this card supports triple-monitor setups for maximum productivity.
- 【Compact Size & Wide Compatibility】: Measuring 240x135x45mm (9.45x5.31x1.77 inches), this dual-fan RX 580 fits perfectly into standard ATX Mid-Towers, Micro-ATX (M-ATX), ideal for compact desktop PC upgrades and space-saving gaming builds.
- 【Optimized Gaming Performance】: With 2048 Stream Processors and a 1206 MHz core clock, this card delivers solid frame rates in popular titles like Fortnite, GTA V, Apex Legends, and Valorant. It’s the ideal budget-friendly GPU for entry-level to mid-range gaming rigs.
- 【Advanced Thermal Management】: Engineered with a dual-fan cooling system and high-efficiency heat pipes to ensure stable performance under heavy loads. The intelligent fan control keeps your system quiet during light office work and provides maximum airflow during intense gaming sessions.
- 【Ready for Content Creation】: Supports DirectX 12, Vulkan, and OpenGL 4.6, making it more than just a gaming card. It provides hardware acceleration for video editing in Premiere Pro, 3D rendering in Blender, and smooth streaming for aspiring creators.
Understand the SD3-specific controls
SD3 exposes controls that are easy to treat as magic numbers. Use them as experiment variables and record them with every useful image.
| Control | Practical starting point | What changing it means |
|---|---|---|
| SD3 TextEncs | CLIP-only on a limited-resource system | SwarmUI exposes CLIP, T5, or both. The project documents CLIP-only as the default because it is generally much faster on constrained systems, with near-identical results in its documented comparison. That is an operational recommendation, not a guarantee that all prompts and images will match. |
| Sigma Shift | 3 | Three is the documented SD3 default. The project notes that experimenting with a lower value around 1.5 can be reasonable, but changing it excessively without a specific reason makes comparisons harder. |
| Steps | 28 as a model-card example | More steps are not automatically better. The model card uses 28 steps in its Diffusers example; treat that as a reproducible starting point, not a universal SwarmUI optimum. |
| Guidance scale | 7 as a model-card example | Guidance influences how strongly the generation follows conditioning. The model card’s value of 7 is an example baseline, not a promise that every subject benefits from it. |
| Refiner Do Tiling | Enable for SD3 refinement or upscaling | SwarmUI recommends tiled refinement/upscaling for SD3 because regular upscaling can produce poor results with this model family. |
Sampler, resolution, aspect ratio, seed, and batch size also matter. There is no single combination that is best for every subject or graphics card. Start with the current default sampler, a resolution your hardware can actually sustain, the model-card example settings where appropriate, and a fixed seed. Change one major variable at a time.
A disciplined first-generation workflow
1. Establish a baseline
Choose SD3 Medium, keep the default Sigma Shift at 3, select CLIP-only if memory is tight, and use 28 steps with guidance 7 as an illustrative baseline from the model card. Keep the sampler, resolution, aspect ratio, and seed fixed for the first comparison set.
Free tools Windows power users keep installed
One-click scans. No signup required.
Use a small batch rather than spending all available time on one image. The point of the first batch is to see how the prompt, composition, and hardware behave—not to prove that one setting is perfect.
2. Write a structured prompt
A reliable starting prompt identifies the:
- Subject: what the image must depict.
- Medium or visual language: for example, oil painting, watercolor, editorial still life, or cinematic concept art.
- Environment: location, surface, background, season, or atmosphere.
- Lighting: direction, softness, color, or time of day.
- Composition: close-up, wide view, centered subject, asymmetrical balance, foreground and background relationships.
- Format: portrait, landscape, square, or another intended aspect ratio.
Illustrative prompt: “A solitary red ceramic vase with three pale branches on a rough wooden table, quiet rural studio, warm side light from a high window, visible brush texture, restrained earth-tone palette, balanced editorial still life, portrait composition.”
This is an example of organization, not a claim that it was personally tested or that it is the best prompt for SD3. Avoid changing the subject, style, lighting, resolution, and sampler simultaneously: if the output changes, you will not know why.
3. Use image history and metadata as your experiment log
SwarmUI’s image history and grid tools are central to the workflow. For every image worth keeping, record or preserve:
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →- the exact prompt;
- seed;
- model and model version or file;
- resolution and aspect ratio;
- sampler;
- step count;
- guidance value;
- SD3 TextEncs selection;
- Sigma Shift;
- any LoRA, ControlNet, refiner, or upscaling settings.
Generate a grid while changing only one factor—for example, compare CLIP-only with both text encoders, or compare Sigma Shift 3 with approximately 1.5. Then keep the stronger result and continue from its seed or metadata. This is more dependable than relying on memory or declaring a setting “best” after one image.
4. Refine only after the composition works
Do not upscale a weak composition in the hope that extra pixels will fix it. First solve the subject, framing, lighting, and major shapes at a manageable base resolution. Then use the refiner or upscaling workflow, with Refiner Do Tiling enabled for SD3 as recommended by SwarmUI’s model-support guidance.
Keep the original base image and its metadata. A refined image is a new workflow result, not a replacement for the reproducible source.
Choosing hardware for local SD3 generation
VRAM is the first question
Raw GPU speed affects how quickly a suitable workload finishes, but VRAM capacity determines whether the pipeline and its working data can fit without aggressive compromises. A fast card with insufficient VRAM can be less useful for this task than a slower card with enough capacity.
For readers buying specifically for local work, an NVIDIA GeForce RTX 4090 graphics card is one recognizable example of a 24 GB graphics-card category that can make local SD3 experimentation more practical. It is not the only suitable option, it is not necessarily the newest or best-value option in every market, and this guide makes no price, availability, or speed claim.
If your GPU has less than 24 GB, try the least disruptive memory-saving changes first:
- Use CLIP-only under SD3 TextEncs.
- Reduce the base resolution or batch size.
- Close other GPU-using applications.
- Use supported offloading or quantization options where your backend and build provide them.
- Accept slower generation before changing many dependencies.
These measures can make a constrained system usable, but no setting guarantees comfortable SD3 Medium generation on a particular card.
NVIDIA TensorRT: faster path, narrower workflow
Stability AI has described collaboration with NVIDIA on TensorRT optimizations for SD3 Medium. SwarmUI separately documents TensorRT support as an NVIDIA-specific accelerator option. The trade-off is flexibility: the TensorRT path is not compatible with some features, including LoRAs and ControlNets, according to the project documentation. Use it when its supported feature set matches your workflow, not as a default replacement for the general pipeline.
Storage, RAM, and the rest of the workstation
A high-capacity NVMe SSD can make it easier to keep model weights, encoders, caches, source images, grids, and output versions on one fast local drive. It does not solve VRAM shortages. Likewise, a desktop RAM upgrade can help when SwarmUI runs alongside Photoshop, a browser, video software, or offloaded model components, but the supplied SD3 documentation does not establish a universal system-RAM threshold.
For a workstation build, also account for power delivery, cooling, and sustained load. These affect stability, but they do not change the model’s licensing or memory requirements.
Rank #3
- All-aluminum metal material - Provides strong and long-lasting support. This is made of all-aluminum metal instead of plastic, can avoid the aging of plastic materials and can be used as a long-term replacement.
- Screw adjustment design - The graphics card bracket design can be compatible with various chassis configurations of traditional and long power supply bays to meet various user hosts.
- Bottom hidden mag.net design - The mag.net hidden in the base is designed for easy installation and more stable standing in the chassis.
- The workmanship of the detail process - The small graphics card support frame is made of three complex processes: polished anode, sandblasted anode and CNC high-speed edge-washing high-gloss process. The full anode process can maintain the durability.
- Tool-free fixing module - The support module is equipped with a cushioning anti-scratch pad and a base high-gloss process.
Troubleshooting SwarmUI and SD3
“CUDA is not enabled” or the GPU is not being used
SwarmUI’s troubleshooting documentation identifies damaged or mismatched PyTorch dependencies as a common cause. Use SwarmUI’s supported repair or reinstall process for the appropriate Torch build instead of randomly installing packages globally with pip. Also check the NVIDIA driver and the startup log for the actual backend error.
Changing several CUDA, Python, and ComfyUI packages at once makes the failure harder to diagnose. Restore the supported environment first; optimize only after a clean generation works.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteAMD on Windows behaves poorly
ROCm support and performance are fast-moving. The project warns that AMD ROCm on Windows has historically been limited and may perform better under Linux. Check the current troubleshooting and platform documentation for your exact GPU, operating system, and backend rather than treating old forum advice as a permanent compatibility rule.
Git reports an ownership or repository error
On Windows, check whether SwarmUI was installed on an external, unusual, or otherwise unsuitable drive. The project recommends a standard internal NTFS drive. Reinstalling dependencies will not necessarily fix a filesystem or ownership problem.
ComfyUI or SwarmUI broke after adding nodes
Undo recent custom-node and package changes where possible. The project warns that indiscriminate use of Comfy Manager or manual package installation can corrupt the backend environment. Keep a clean baseline, add one extension at a time, and test a simple SD3 generation after each change.
The SD3 model does not appear in the selector
- Confirm that the model download completed.
- Confirm that the active backend is the backend whose model directory contains the files.
- Review SwarmUI’s model-root and folder-mapping configuration for existing Auto1111 or ComfyUI directories.
- Restart or rescan using the options available in your current build.
If the model is visible but the first generation fails, check whether the required text encoders are still downloading or whether the error is actually a VRAM or Torch problem.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →The process runs out of memory
Reduce batch size and resolution, select CLIP-only text encoding, close other GPU workloads, and use supported offloading or quantization. If the model loads but generation is extremely slow, that is a different problem from failure to load: offloading may have made the pipeline fit while moving work between system memory and the GPU.
Upscaling creates ugly or unstable results
Return to the base image and verify the composition first. For SD3, enable Refiner Do Tiling before trying the refinement or upscaling pass. SwarmUI specifically recommends tiled refinement because regular upscaling can produce poor results with this model family.
Commercial use, publishing, and record-keeping
Downloading SD3 Medium does not automatically grant unrestricted commercial rights. Stability AI released SD3 Medium with a Stability Community License; its announcement directs large-scale commercial users to obtain appropriate licensing from Stability AI. Stability’s license-update page describes free commercial use for individuals and small businesses below a stated annual-revenue threshold, but the exact threshold, conditions, and model-specific terms must be checked in the current license before publication or deployment.
Before selling images, using them in a client project, or building a commercial service, verify:
- the current SD3 Medium license and any applicable commercial-use threshold;
- the terms for every LoRA, ControlNet, VAE, model, embedding, or other asset in the workflow;
- the license for datasets or reference material used in training or production;
- the terms of the software and backend, including any extensions;
- the rules of the platform where the work will be published or sold;
- privacy, publicity, trademark, copyright, and other rights relevant to the prompt and output.
Keep a simple production record containing model filenames and versions, downloaded licenses, LoRA and ControlNet sources, dataset notes, prompts, seeds, and software versions. Generated images are not automatically free from legal or platform restrictions merely because they were produced locally.
Recommended SD3-in-SwarmUI checklist
- Use the official SwarmUI repository and current documentation.
- Install on a supported Windows, Linux, or Apple M-series system.
- On Linux, use Python 3.10–3.12 and avoid Python 3.13 unless the project later documents support.
- Choose ComfyUI as the normal local backend.
- Keep a private access mode for a single-user installation.
- Use authentication and HTTPS before public hosting.
- Confirm the model root and folder mapping if SD3 Medium is not visible.
- Allow the first-run SD3 text encoders to download.
- Start with CLIP-only if VRAM or speed is limited.
- Use Sigma Shift 3 initially; experiment cautiously around 1.5 rather than changing it arbitrarily.
- Treat 28 steps and guidance 7 as model-card starting values, not universal defaults.
- Record seed, sampler, resolution, steps, guidance, text encoders, and Sigma Shift.
- Enable Refiner Do Tiling for SD3 refinement or upscaling.
- Keep the backend environment clean and avoid indiscriminate package or custom-node changes.
- Review licensing before commercial publication.
SwarmUI is most valuable when it turns SD3 experimentation into a repeatable process: a consistent model and backend, controlled parameter changes, searchable history, and a refinement stage that respects the model’s memory and upscaling behavior. The interface can make that process easier, but the quality of the final art still comes from informed choices throughout the pipeline.
Frequently Asked Questions
Is SwarmUI itself a Stable Diffusion 3 model?
No. Stable Diffusion 3 is the model family, while SwarmUI is a web interface and workflow manager. A typical local SwarmUI installation uses ComfyUI as its backend, and Stable Diffusion 3 Medium supplies the model weights.
Can Stable Diffusion 3 Medium run on a 12 GB or 16 GB GPU?
It may be possible, but the official Diffusers documentation warns that SD3 can be challenging below 24 GB of VRAM even with FP16. Lower-VRAM systems may need CLIP-only text encoding, reduced resolution or batch size, offloading, quantization, or a slower workflow.
Recommended Free Tools
What are good starting settings for SD3 Medium in SwarmUI?
Use Sigma Shift 3 initially, and treat 28 steps and guidance scale 7 as example starting values from the SD3 Medium model card. SwarmUI documents CLIP-only text encoding as generally faster on limited-resource systems, and recommends tiled refinement for SD3. None of these settings is universally optimal.
Can I sell images made with Stable Diffusion 3 Medium?
Not automatically. SD3 Medium is released under a Stability Community License, and commercial-use conditions can depend on the current license, business size, revenue threshold, and scale of deployment. Check Stability AI’s current license terms and the licenses for all additional models and assets before commercial use.
The Bottom Line
Bottom line: Use SwarmUI as the workflow layer and Stable Diffusion 3 Medium as the model, with ComfyUI handling the local backend. Begin conservatively, treat 24 GB of VRAM as a practical target rather than a mandatory specification, compare settings systematically, enable tiled refinement, and verify licenses before publishing or selling the results.
Quick Recap
