Stable Diffusion

The most reliable structure is: subject → style/medium → composition/lighting → quality boosters. Example order: "a red vintage bicycle, oil painting, golden hour lighting, close-up, highly detailed." Front-load the subject — tokens near the start of the prompt carry more weight in the final image.

A negative prompt tells the model what to avoid, without spending tokens describing it positively. Standard starting negative prompt: "blurry, low quality, extra limbs, deformed hands, watermark, text, bad anatomy." Negative prompts are the single fastest fix for the classic Stable Diffusion hand and limb problems.

Wrap a term in parentheses to increase its weight — (neon lighting:1.4) pushes that concept harder than the default weight of 1.0. Square brackets reduce weight the same way in reverse. Weighting is how you resolve conflicts when the model is ignoring part of a longer prompt.

Overloading a prompt with 40+ competing keywords dilutes all of them. Vague adjectives like "beautiful" or "amazing" do almost nothing — the model needs concrete visual descriptors. Forgetting a negative prompt entirely is the most common reason first-session AI art looks unfinished.

Take a photograph and render it as oil painting, watercolour, anime, or any other visual style via prompt. The composition and subject are preserved while the rendering style changes entirely. A core use case for social content and commercial creative work.

Change specific elements of an existing photo — lighting, environment, clothing, background — without rebuilding the entire image. Combined with inpainting, you can select specific regions for targeted AI transformation while leaving the rest untouched.

Take a raw product photograph and generate variations — different environments, lighting conditions, colour options — without reshooting. E-commerce teams use img2img to produce entire product photography sets from a single hero shot.

Upload a rough sketch or wireframe and prompt Stable Diffusion to render it as a finished illustration, product mockup, or architectural visualisation. The spatial structure of the sketch is preserved; the rendering quality is added by the model.

Mask a specific region of an image and generate new content for that region only — remove objects, replace backgrounds, add or change elements. Inpainting is available directly in A1111's img2img tab and is one of the most practical AI photo editing tools available.

Generate a rough direction from text, then use img2img to refine, iterate, and steer toward the final result — a faster workflow than starting from scratch each iteration. Used by concept artists, game designers, and filmmakers for rapid visual development.

Trains a small set of additional weights (typically 20–200MB) layered on top of the base model. Fastest to train, easiest to share, and the default choice for most product, character, and style fine-tunes in 2026.

Fine-tunes the full model on your subject, producing the strongest and most faithful reproduction of it. Slower to train and produces a full model checkpoint (2–7GB) rather than a lightweight add-on file.

Trains a new embedding — a single new "word" the model learns to associate with your concept — without modifying the model weights at all. Smallest file size, fastest to train, but the least flexible of the three.

LoRA for most product, brand, and character work. DreamBooth when you need maximum fidelity and can tolerate a longer training run. Textual Inversion for quick style or concept experiments that don't need to travel between models.

The oldest actively-used version. Requires only 4GB VRAM — runs on budget hardware. The Civitai ecosystem has more fine-tuned models built on SD 1.5 than any other base. Output quality is lower than newer models but the flexibility and model availability are unmatched. Still the right choice for constrained hardware.

Significantly better output quality than SD 1.5 — higher resolution, better prompt following, stronger photorealism. Requires 8GB VRAM. A two-stage pipeline (base + refiner) produces the best results. The current standard for most production workflows that don't need Flux-level quality.

Improved text rendering and compositional accuracy over SDXL. Requires 10–12GB VRAM. The licence differs from earlier versions — verify commercial use terms before deploying. A strong choice for use cases requiring accurate text rendering inside images.

Flux from Black Forest Labs (original SD research team) represents the current state of the art for open-source generation. Dramatically better image quality, photorealism, and prompt adherence than any previous version. Requires 12–24GB VRAM — effectively a high-end desktop GPU or cloud instance. Preferred via ComfyUI for best results. The model serious local generation workflows are converging on in 2026.

A browser-based, fully hosted AUTOMATIC1111 or ComfyUI instance. No installation at all — you get the same interface you'd run locally, streamed to your browser, with the GPU handled entirely on RunDiffusion's side.

Free (and paid) GPU-backed notebooks. Widely used for one-off generation sessions and running community A1111/ComfyUI notebooks without owning a GPU at all — the free tier has session time limits and queue waits.

A GPU rental marketplace — rent a specific GPU (including high-end cards for Flux) by the hour from a global pool of providers, then install and run your own stack exactly as you would locally.

RunDiffusion for a plug-and-play experience with zero setup. Colab for occasional, casual generation. Vast.ai when you need a specific high-end GPU on demand without owning one.

No content policy you did not write. No filter between your prompt and the output. For commercial studios, product photography pipelines, and art direction work — the absence of a platform moderating your output is a meaningful capability difference.

Structural conditioning via pose, depth, edge, and line art maps. Turns generation from probabilistic to directed. The capability gap between Stable Diffusion with ControlNet and any closed platform is significant and not closing.

Transform existing photos, sketches, and renders via text prompt. Style transfer, AI photo editing, inpainting, product visualisation — capabilities that go far beyond text-to-image. The img2img workflow is unique to locally-run open-source generation at this depth of control.

Train on your own images with Dreambooth or LoRA. 10–30 reference photos. The model learns your product, your face, your brand aesthetic. No enterprise contract required. No data leaving your machine.

AUTOMATIC1111 and ComfyUI both expose REST APIs. Run headless, integrate into product workflows, trigger batch generation pipelines, or build your own generation interface on top. Closed platforms do not allow this depth of integration.

Civitai hosts thousands of community-trained models — anime, photorealism, architecture, fashion, product photography styles. Switch models for different visual domains without retraining from scratch. The largest open model ecosystem in this category.

There is no hosted experience. Python, CUDA, and 4–24GB VRAM depending on your target model. If this is a barrier, use a hosted wrapper like RunDiffusion or Google Colab — you trade control for ease. The local path rewards those who clear the setup wall.

SD 1.5 runs on 4GB VRAM — a budget GPU. SDXL needs 8GB. SD3 needs 10–12GB. Flux requires 12–24GB — effectively a high-end desktop or cloud GPU. A setup that worked for SD 1.5 will hit a wall with Flux. Check your hardware before choosing a model version.

The same prompt produces wildly different results across model versions, samplers, and CFG settings. There is no "just type what you want." Output quality is proportional to prompt literacy. Expect a deliberate learning investment before outputs feel reliable.

Without fine-tuning or ControlNet conditioning, generating consistent characters, products, or environments across multiple images requires significant workflow engineering. This is one of the harder problems in the entire AI image category — and Stable Diffusion solves it better than any closed tool.

CreativeML Open RAIL-M permits commercial use of the model — but does not resolve questions about training data copyright, the copyrightability of AI-generated output, or your liability when fine-tuning on third-party images. Commercial use warrants independent legal assessment.

Official docs are sparse. The real knowledge base lives in Reddit communities, YouTube tutorials, Civitai model pages, and Discord servers. Learning Stable Diffusion means engaging with that ecosystem — it is genuinely one of the best technical communities in AI.

The licence covering SD 1.5 and SDXL. It permits commercial use of the model and its outputs, with a short list of restricted use cases (e.g. generating illegal content) baked into the licence text itself.

SD3 and SD3.5 ship under different terms than SD 1.5/SDXL — commercial use above certain revenue thresholds may require a separate licence from Stability AI. Always check the current terms before deploying SD3-family models in a commercial product.

Generated output can generally be sold or used commercially under the applicable model licence — but the licence covers your right to use the model, not a guarantee that the output itself is free of third-party claims.

Stability AI has faced litigation from Getty Images and a group of visual artists over the use of copyrighted material in training data. These cases are ongoing and could affect future licensing terms — worth monitoring if your use case is commercially significant.

The open licence does not resolve whether the underlying training data was used lawfully in every jurisdiction. This question sits upstream of anything you generate and is unrelated to how you personally use the model.

Whether a purely AI-generated image can be copyrighted at all — and by whom — remains unsettled in several jurisdictions, including notably the US Copyright Office's evolving guidance. Human-edited or human-directed compositions tend to fare better than a single unedited generation.

A traditional shoot means booking a studio, a photographer, and a retouching pass — typically a multi-day cycle per product line. Stable Diffusion collapses that into a same-day workflow once a base product shot and a trained LoRA exist.

A local GPU suits ongoing, high-volume catalogue work where the hardware pays for itself over time. A cloud option like RunDiffusion or Vast.ai suits occasional campaigns where owning a GPU year-round doesn't make sense.

A single generation takes seconds. A full reshoot cycle for a new background or lighting concept in a traditional studio can take one to two weeks. That speed gap compounds fast across a large catalogue or a fast-moving campaign calendar.

Once a product LoRA exists, generating variations — new backgrounds, new lighting, new seasonal themes — scales far more easily than rebooking physical shoots for every SKU and every campaign refresh.

The real overhead isn't the generation itself — it's the setup time, the prompt engineering learning curve, and the LoRA training pass needed to keep every generated shot consistent with the actual product.

A well-trained product LoRA keeps the product itself pixel-consistent across hundreds of generated variations, which is a harder problem to solve consistently with human retouchers across a large team.

An AUTOMATIC1111 extension that generates animated sequences by interpolating prompts and camera movement across frames — the standard entry point for AI-generated animation built on Stable Diffusion.

Stability AI's dedicated image-to-video model, purpose-built to animate a still image into a short video clip with coherent motion, rather than repurposing the image model frame by frame.

The common pattern: generate a strong still image first with full prompt and ControlNet control, then feed that single frame into Stable Video Diffusion to produce motion — combining the precision of image generation with the emerging capability of video generation.

Video generation is significantly more VRAM- and time-intensive than a single still image — expect to need the same high-end GPU tier as Flux, and generation times measured in minutes rather than seconds per clip.