Compare SD 3.5 models and learn local workflows, ComfyUI, LoRA, ControlNet, reproducibility, hosted access, and commercial licensing.
Stable Diffusion gives creators access to downloadable image-model weights and customizable workflows. You can choose how a model runs, add compatible adapters, and process assets on your own infrastructure instead of relying entirely on a fixed app.
The benefit is control; the cost is setup and maintenance. This guide covers SD 3.5, local tools, structural guidance, reproducibility, and commercial licensing. Visit Stability AI for official releases.
Stable Diffusion is a family of image-generation models associated with Stability AI. Many releases offer downloadable weights, but that does not mean every version has the same license or works with the same extensions.
As checked October 8, 2026, Stability’s Core Models catalog lists SD 3.5 Medium, Large, and Large Turbo. A newer hosted image service should not automatically be described as a new downloadable Stable Diffusion checkpoint.
You can run compatible models locally with ComfyUI or Diffusers, access hosted services, or use third-party applications. Choose the route according to your privacy, customization, and maintenance requirements.
The SD 3.5 release announcement introduced Large and Large Turbo in October 2024, with Medium following on October 29.
| Model | Parameters | Practical role |
|---|---|---|
| SD 3.5 Large | 8.1 billion | Quality-focused generation |
| SD 3.5 Large Turbo | Distilled Large model | Faster generation with four inference steps |
| SD 3.5 Medium | 2.5 billion | Lighter local experimentation |
Large targets one-megapixel generation, while Medium supports roughly 0.25 to two megapixels according to the release notes.
Four-step Turbo sampling does not guarantee the same quality as a longer Large run. Compare outputs on the actual subjects you need. SDXL can also remain useful when your workflow depends on a particular community checkpoint or adapter; compatibility may matter more than choosing the newest family.
Use the official SD 3.5 repository and a workflow intended for your exact checkpoint. A model file alone may not include all required text encoders or other components.
For a visual workflow, ComfyUI exposes generation as connected nodes. For programmatic work, Diffusers provides model pipelines. Begin with a basic generation example before adding custom nodes, adapters, or batch automation.
Hardware requirements depend on precision, quantization, resolution, batch size, and component offloading. Avoid a universal VRAM minimum. Offloading can lower GPU memory use while increasing system-memory needs and latency.
Record a successful baseline: checkpoint, encoders, image size, sampler or scheduler, steps, guidance, and software versions. That gives you something to return to when an extension changes behavior.
LoRA adapters modify a model’s behavior with smaller trained weights, often for a subject or style. ControlNet adds structural conditioning from inputs such as edges or depth.
Stability’s official SD 3.5 Large ControlNets include Canny, Depth, and Blur. These are trained for Large; do not assume an SDXL adapter or a Large ControlNet will work unchanged with Medium.
A useful workflow is to establish composition first, then refine style and detail. For example, use a compatible depth guide to preserve a product arrangement, then test whether a LoRA improves the desired look without distorting the subject.
Keeping the seed fixed helps compare changes within a controlled setup. However, Diffusers’ reproducibility guide explains why random-number generators and execution conditions matter.
The same seed is not a promise of identical pixels across different hardware, library versions, or pipelines. Save the workflow and environment information alongside the seed, and reset the generator appropriately when comparing runs.
For client work, retain the original references and final exports as well as generation settings. Reproducibility is most useful when it helps you revise an approved result reliably.
The Stability AI license page states that covered Core Models can be used commercially without a license fee by individuals or organizations below $1 million in total annual revenue, subject to the Community License.
The threshold concerns total revenue, not just money earned from generated images. Stability directs organizations reaching the threshold to discuss an Enterprise License. Purely non-commercial research is treated separately.
Check the exact checkpoint and derivative license. Community LoRAs, fine-tunes, and other components may introduce additional conditions. Downloadable weights are not automatically unrestricted open-source software, and permission to use a model does not guarantee legal clearance for every image.
In an October 2024 comparison, The Walking Fish published SD 3.5 Large, Turbo, and FLUX.1 dev outputs from matching prompts.
The author found a convincing Large portrait in one example, but reported an extra finger and awkward umbrella structure in another. These examples show why attractive previews still need detail inspection.
This is an early, limited comparison rather than a current leaderboard. It does not establish how every fine-tune performs, and this guide does not claim firsthand testing.
> A vintage leather backpack on a mossy stone wall, golden-hour light, product photography, shallow depth of field, rich leather texture, uncluttered background.
For hosted production, consult Stability’s Stable Image documentation. Confirm the endpoint, capabilities, pricing, and terms rather than assuming a hosted service reproduces your local checkpoint.
Stable Diffusion is useful when customization and infrastructure control justify the extra work. Build one dependable workflow first, then add adapters and automation as your project requires.
Stable Diffusion is a family of image-generation models. Many releases offer downloadable weights for local workflows, while hosted services and third-party applications provide other access routes.
Large prioritizes quality, Large Turbo uses four-step distilled generation, and Medium is a lighter local option. Compare them using your actual brief and compatible workflow.
Requirements depend on the checkpoint, precision, quantization, image size, batch size, and offloading. There is no universal VRAM minimum for every workflow.
No. LoRA adapters modify model behavior, often for a style or subject. ControlNet supplies structural guidance such as edges or depth. Both must be compatible with the selected model.
No. A fixed seed helps controlled comparisons, but hardware, software versions, random-number generators, and pipeline settings can change results. Save the complete workflow and environment.
Covered models allow commercial use under the Community License for individuals or organizations below $1 million in total annual revenue, subject to its terms. Review enterprise requirements when reaching that threshold and check all derivative licenses.
Eligible model use may have no license fee, but local inference still requires hardware, electricity, storage, and maintenance. Hosted services have their own costs and terms.
Yes. Its developer platform provides Stable Image services. Check the specific endpoint and terms rather than assuming it matches a local SD 3.5 checkpoint.
