Stable Diffusion 3.5 in Practice: Three Models and Editing Tools

Stable Diffusion is a model family bundled with editing tools

Stable Diffusion is Stability AI's family of image generation models. The same page presents it under the name Stable Image, together with the editing, upscaling and control tools from the same company. The official product page calls Stable Diffusion 3.5 the company's "most powerful image model yet" and splits it into three models: Large, Turbo and Medium. The previous generation, Stable Diffusion XL, and its distilled sibling XL Turbo are listed as additional models.

The point that matters for practitioners is that this is not a single app. The same models can run on your own server, plug in through an API, or be used directly in a web-based creative tool. Each route changes your cost structure and how much control you have, so pick the deployment path before you pick a model name. It saves a lot of trial and error.

Why mentions clustered in a short window

The Muckle radar gives this topic a signal score of 92.6/100, which puts it in the "overheated" stage. Mentions over the last 7 days came to 20, up sharply from 3 in the 7 days before, and the 30-day total is 23. Six channels and 19 people or sites discussed it, up from 3 the week before. The first detection was on 2024-07-04. The absolute counts are small, so read this less as a mass-market burst and more as several conversations landing in the same week. The "overheated" label reflects speed of growth, not scale.

Grouping the recent posts by theme in our own analysis shows a few clusters. The first is local execution: whether to buy or rent a graphics card, how far an older GPU can go, and whether video is feasible locally. The second is tooling: keeping characters consistent with LoRA training, using the models inside node-based workflow tools, and how negative prompts actually work. The third is ethics and rights, including who owns a generated image and the underrepresentation of minorities. The newest lineup confirmed on the official page is the 3.5 family, so there is no basis in official documentation for tying this spike to a specific new release. The safer reading is that existing practical discussions piled up in one week.

Three models and the editing tools are the core

The official page names three strengths for 3.5. First is a wide range of styles, including 3D, photography, painting and line art. Second is prompt adherence that the company says is comparable to much larger models. Third is output diversity that reflects people and scenes from around the world. The prompt-adherence claim is the vendor's own description, and the page gives no numeric comparison.

Here is how the models differ.

The editing and control tools are packaged separately from generation. On the editing side you get erase object, inpaint, outpaint, remove background, search and recolor, search and replace, and background replacement with relighting. Upscaling comes in three modes: Creative, Conservative and Fast. The control tools are sketch, structure and style. The page says all of them work on both uploaded and generated images.

For a sense of how the underlying technology works, public research is a useful reference. One study explains that Stable Diffusion learns diffusion in a latent space produced through a VQGAN rather than in pixel space, which helps efficiency and quality and also supports inpainting and local edits. Another description says text prompts are encoded with CLIP. Both accounts describe older versions covered by the research, and neither confirms the internal architecture of 3.5.

In design, video and 3D, tool combinations matter more than raw generation

Design drafts. Suppose you are building a brand campaign. Start with Turbo to pull many quick variations of composition and mood and narrow the direction. Once you have one, push the same prompt through Large for final candidates. If you have a rough sketch, use sketch control to lock the composition while the model adds texture and color. Product color variants can go through search and recolor, and prop swaps through search and replace, so you never regenerate the whole image. Finish with Conservative upscaling, which raises resolution while keeping the original look.

Video pre-production. The official page describes 3.5 as an image model, so do not treat it as a tool that makes video. For video teams the realistic use is storyboards and keyframes. Build a reference image for a scene, apply style control to give other shots the same look, and use background replacement with relighting to place the same character at different times of day. You can hand those keyframes to a video generation tool as starting images, but quality at that stage depends on the separate tool. Character consistency across cuts is where extra work such as LoRA training comes in, and it is a recurring complaint in practitioner reports we reviewed.

3D support. The official description includes 3D-style generation, so the models suit concept art and reference images before 3D production. You can produce front and side character concepts, isolate them cleanly with background removal for a modeler, or use structure control to match the form of a texture reference. The official page does not describe direct 3D mesh generation, though, so the step from image to 3D needs another tool.

Developer integration. If you are adding image generation or editing to an app, the API is the fastest route. Inpainting, upscaling and background removal are called per endpoint, which fits services that process photos users upload. If you must run on your own infrastructure, the self-hosted license lets you deploy and layer on fine-tuning or LoRA.

Pick alternatives within the family, and test outside models yourself

As the table shows, choices inside the family come down to a balance of speed, quality and hardware demands. Speed research is worth a look too. The LCM-LoRA study describes a distillation module applied to SD v1.5, SSD-1B and SDXL that can be attached to various fine-tuned models or LoRAs without extra training. The figure of about 32 A100 GPU hours cited there is the cost of distillation training, not an inference requirement for users. Another study reports generation speedups of 22.4% and 40.0% from model assembly and feature inheritance, and a study on arbitrary-size generation reports inference time cut in half compared with an existing tiling method. Each number comes from that study's experimental conditions and cannot be converted into speed on your machine.

This report does not rank the family against other companies' hosted image services or other open models. We found no recent evaluation, run with identical prompts, seeds, resolutions and editing conditions, that we could rely on. You can still set clear selection criteria. If you want to host, control and customize on your own server, this family's self-hosting option is an advantage. If you want to avoid setup and operations entirely, a web app or API service is the better fit. In both cases, the most reliable method is to compare candidates yourself on around ten of your most-used prompts.

Check licensing, cost, requirements, bias and safety first

License and cost. The official page puts a "Get license" path on every model and lists self-hosted licenses, the API and cloud partners as separate routes. It does not state pricing, the scope of commercial use or redistribution terms. For commercial projects, read the license document for the exact model and route first, and confirm that your delivery or service scope is allowed.

Hardware requirements. The official description tells you only that Medium is designed for consumer hardware and that Large targets professional use at 1 megapixel resolution. Minimum VRAM per model is not on the page. Do not turn distillation training costs, or memory-saving claims with no quantitative conditions, into minimum specs. If you plan to run locally, check the specs on the official model card and test on your own GPU across resolutions and sampling settings.

Bias. Public research repeatedly reports representational bias in image generators. One analysis of more than 96,000 generated images found that three systems, including Stable Diffusion v1.4 and v2, all over-represented whiteness and masculinity. Another evaluation study edits images of people with neutral prompts to see shifts in gender, race and age. The 90.8% that study reports is the accuracy of its automatic bias detection, not a bias rate for Stable Diffusion. All of these results are limited to older versions and specific research designs, so they do not transfer directly to 3.5. Because the official page lists output diversity as a strength, though, anyone producing people imagery at volume should audit results directly.

Safety filters. A 2022 study reported that the safety filter it tested could be bypassed and did not cover content such as violence and gore. A 2023 study reported that about half of the prompts classified as safe could be manipulated to get around several safeguards. These too are results for past mechanisms. Developers wiring the models into a service should add their own review step on both input and output, and test the safety features of the exact version in use.

Reviewing edits. According to research, the base VQGAN can distort regions that were not edited, and an asymmetric decoder has been proposed to reduce this. Whether that improvement is in the current product is unconfirmed. Make a habit of zooming in on both the edited area and the untouched original area around it when you compare partial-edit results.

Getting started depends on your goal, and there are four routes

The lightest way in is Stable Assistant, the web-based creative platform. It puts every model and editing tool in one place, which makes it a good spot to build prompt intuition and judge which tools fit your work. App developers can start with the API and add one feature at a time, such as inpainting or background removal. With a cloud partner, you can call the models from a cloud environment you already use, with no separate infrastructure. If you want to run locally or add training, getting a self-hosted license and starting with Medium is the realistic choice.

A good first-week exercise is to run the same prompt through Turbo, Medium and Large and record the differences in speed and quality. Then pick one or two edits you use often and measure how much repetitive work they cut. Check the license terms at this stage too.

Muckle's verdict

For teams that repeat image-centered work such as design drafts, storyboards and concept art, the combination of the three 3.5 models and the editing tools is well worth testing. Offering both self-hosting and an API is an advantage for teams that care about control. Still, license terms, per-model hardware requirements and the scope of current safety features remain items to confirm in official documentation, so settle those three before any commercial rollout. This spike in mentions is not tied to a specific new announcement, so keep an eye on when the official lineup changes.