Stable Diffusion on Amazon Bedrock: models, pricing and prompting.

IStability AI's image models on Amazon Bedrock

Stability AI — the company behind Stable Diffusion — is one of the foundation-model providers on Amazon Bedrock, supplying image-generation models through Bedrock's single managed API alongside Amazon's own Nova and Titan, Anthropic Claude, Meta Llama, Mistral, Cohere, and others. On Bedrock the Stability lineup spans two generations, and picking the right model is the central cost-and-quality decision.

As of 2026 the Stability models on Bedrock fall into two groups. The first is the SDXL-era Stable Diffusion family — the SDXL 1.0-class text-to-image model that put Stable Diffusion on Bedrock originally, billed by step count and resolution and still useful for cost-sensitive, controllable generation. The second, and where most new work should start, is the newer Stable Image generation line: Stable Image Core (fast and low-cost, tuned for speed and reliable prompt adherence without heavy prompt engineering), Stable Diffusion 3.x (the SD3 / SD3.5-class models with markedly better prompt following, multi-subject composition, and typography), and Stable Image Ultra (the highest-quality tier, built on the most advanced Stable Diffusion architecture for photorealism and fine detail).

The practical discipline mirrors the one that governs all Bedrock cost: match the model to the task. Use Core for high-volume drafts, thumbnails, and iteration where speed and price matter; use the SD3.x tier as a strong general default with good prompt adherence; reserve Ultra for hero images and final assets where quality dominates and the higher per-image price is justified. Because every model sits behind the same Bedrock invoke path, moving a generation request from one Stability model to another is essentially a change of model ID — which makes a draft-on-Core, finish-on-Ultra workflow easy to build.

Image models differ from text models in one structural way worth stating up front: they are billed per generated image, not per token. There is no input/output token meter; the unit of cost is one image at a given model and quality tier (and, for the SDXL-era model, at a given step count and resolution). That changes how you budget and how you optimize — the levers are model choice, resolution, step count, and how many candidate images you generate per prompt, not context length or caching.

One caveat, stated once and meant throughout: exact model names, model IDs, regional availability, supported resolutions, and per-image prices all change as Stability ships new models and AWS updates Bedrock. The names, identifiers, and figures here are representative as of 2026 to convey the structure and relative cost. Always confirm the current model IDs in the Bedrock model catalog and current rates on the AWS Bedrock pricing page before you build or budget.

IIModel access and invoking a Stability model

Access to many Bedrock models, including Stability's, is on by default. Stability AI models are sold through AWS Marketplace, so the first call in your account subscribes it in the background; after that you call the model through the Bedrock runtime.

First use. The role that makes the first call needs aws-marketplace:Subscribe, aws-marketplace:Unsubscribe and aws-marketplace:ViewSubscriptions, and the account needs a valid payment method; AWS says the subscription can take up to 15 minutes to finish. After that, other identities only need the Bedrock invoke permission. You pay only when you generate an image. Image-model Region availability tends to be narrower than the big text models, so confirm the model you want is offered in your Region before you design around it. Details are in the Bedrock model access docs.

Model IDs. Every model on Bedrock is invoked by a model ID — a string identifying the provider, model, and version, namespaced under Stability (of the shape stability.… with a version suffix). You pass this ID to the API to choose which Stability model generates an image, so moving a request from Core to SD3.x to Ultra is a change of model-ID string. Because IDs advance as Stability ships new models, do not hard-code a guessed value — read the current ID from the Bedrock model catalog in the console or list it via the API/CLI, and treat it as configuration rather than a literal in your code.

Invoking. Image models are called with Bedrock's InvokeModel action (the model-agnostic Converse API is for conversational text/multimodal models, not image generation — so image generation uses the per-model invoke path with a model-specific request body). You send a JSON body containing the prompt and the model's generation parameters and receive the generated image (typically as a base64-encoded payload) in the response. The exact request schema differs between the SDXL-era model and the newer Stable Image models — the SDXL-era body exposes parameters like cfg_scale, steps, seed, style_preset, and dimensions; the Stable Image models take a simpler body centered on the prompt, an aspect ratio, an optional negative prompt, a seed, and an output format. Read the model's parameter reference for the exact fields.

Permissions. The IAM principal making the call needs permission for the relevant Bedrock invoke action on the specific Stability model ARNs. A least-privilege policy scoped to the model IDs you intend to use is the recommended posture — the same IAM, VPC endpoint, KMS, and CloudTrail controls that govern the rest of your Bedrock usage apply to image generation too, so generated images and prompts stay inside your account and region.

• Make the first call from a role with AWS Marketplace subscribe permissions (and a valid payment method on the account); Bedrock subscribes the model on first use.
• Confirm the model is offered in your region — image-model availability is narrower than the major text models.
• Get the current model ID from the model catalog or via the API — do not hard-code a guessed version string.
• Call InvokeModel with a model-specific JSON body (prompt + generation params); read the base64 image from the response.
• Attach a least-privilege IAM policy for the Bedrock invoke action on the specific Stability model ARNs. You pay only per generated image.

IIIPer-image pricing — how Stability is billed on Bedrock

Image generation on Bedrock is priced per generated image, not per token. The rate depends on the model: the SDXL-era Stable Diffusion model is priced by step count and resolution, while the newer Stable Image models use a flat per-image price by tier (Core cheapest, SD3.x in the middle, Ultra highest).

This is the most important structural difference from the text models elsewhere in this cluster. There is no per-1K-token meter and no prompt caching — the unit of cost is one finished image. For the SDXL-era model, the price scales with the number of diffusion steps and the output resolution: a standard-resolution image at the default step count costs a few cents, and the price rises with higher step counts and larger images (more steps and pixels = more compute = higher per-image cost). For the Stable Image models, the price is a flat per-image figure per model tier regardless of step count, which makes budgeting simpler — you multiply expected images by the tier's per-image rate.

The table below gives representative 2026 per-image rates to rank the tiers and sanity-check a budget — not as an audited price sheet. The dominant cost levers for an image workload are different from a text one: model/tier choice (Core vs SD3.x vs Ultra), resolution and step count (for the SDXL-era model), and how many candidate images you generate per prompt (generating four variations costs four images). Generating a contact sheet of cheap Core drafts and then re-rendering only the chosen composition on Ultra is the per-image analogue of the tiered-routing pattern that governs text-model cost.

IVCapabilities: text-to-image, image-to-image, editing, and controls

Stability's models on Bedrock cover more than a single text prompt in, one image out. The lineup supports generation from text, transformation of an existing image, in-place editing, and a set of controls over composition and style. Exactly which capability a given model exposes varies by model and version, so confirm specifics for the one you choose.

VPrompting Stable Diffusion well

Image prompting is its own craft, and the newer Stable Image models reward clear, structured natural-language prompts more than the keyword-soup style that older Stable Diffusion communities relied on. A few durable practices get you most of the way.

The single biggest shift with the SD3.x / Stable Image generation is that they follow descriptive, sentence-style prompts well — you can write what you want as a clear description rather than stacking comma-separated tags and weights. State the subject, the setting, the composition and framing (close-up, wide shot, overhead), the lighting (soft daylight, dramatic rim light, studio softbox), the style or medium (photograph, 3D render, watercolour, line art), and any mood or colour direction. Specificity beats length: a precise, well-ordered description outperforms a long pile of adjectives.

• Lead with the subject, then qualify it: Put the main subject first and add detail around it — “a ceramic coffee mug on a marble countertop, morning light from the left, shallow depth of field” reads far better than a tag list. The model weights earlier, clearer elements more reliably.
• Describe lighting, framing, and medium explicitly: Lighting and camera framing do more for realism than any “high quality” incantation. Name the shot type (close-up / wide / overhead), the light (soft, hard, golden-hour, studio), and the medium (photo, render, illustration). For product and marketing work this is where usable output comes from.
• Use the negative prompt to subtract, not to add: A negative prompt lists what to exclude (extra fingers, text artifacts, harsh shadows, watermarks). Keep it focused — overstuffing it can wash out the result. It is a clean-up tool, not a second positive prompt.
• Pin a seed when you want consistency: Fix the seed to reproduce a result or to make controlled variations (same seed + tweaked prompt = a near-sibling image). Leave it random when you want fresh candidates. Seeds are how you turn a lucky first draft into a repeatable asset.
• Tune cfg / steps only on the SDXL-era model: On the SDXL-era model, a higher cfg_scale follows the prompt more strictly (too high looks over-cooked) and more steps add detail at higher cost. The Stable Image models do not expose these — you steer them through the prompt and aspect ratio instead. Match your knobs to the model.
• Generate a few candidates, then refine the winner: Produce a small batch, pick the strongest composition, then iterate on it (image-to-image, inpainting, or a pinned seed with prompt tweaks). Because cost is per image, a tight draft-then-refine loop is both cheaper and faster than over-specifying one expensive render.

VIStability vs Amazon Nova Canvas vs Titan Image Generator

Stability is one of several image options on Bedrock. The two others people ask about are both Amazon's own: Nova Canvas (the current Amazon image generator) and the older Titan Image Generator. A neutral orientation — they overlap, and the right pick is workload-specific.

Stability vs Amazon Nova Canvas. Nova Canvas is Amazon's current first-party image model on Bedrock (part of the Nova family), with text-to-image, image variation, inpainting/outpainting, and background removal built in — plus two things some buyers specifically want: an invisible watermark on every generated image and built-in content controls. It is competitive on price and convenient if you are already standardizing on Nova. Stability tends to be the pick when you want the specific Stable Diffusion aesthetic, the breadth of the Stability community's style vocabulary, the SDXL-era model's fine-grained step/cfg control, or the top-end photorealism of Stable Image Ultra. Many teams keep both enabled and route by use case. See the amazon-nova sibling.

Stability vs Amazon Titan Image Generator. Titan Image Generator is Amazon's earlier first-party image model — also text-to-image with editing and an invisible watermark, and in many deployments now superseded by Nova Canvas for new work. It remains a low-cost option and is fine for straightforward generation, but for new pipelines the realistic choice is usually Nova Canvas (Amazon's current model) vs a Stability model rather than Titan. If you are already on Titan, the comparison is whether Stability's aesthetic and Ultra-tier quality, or Nova Canvas's newer feature set, justify a move. See the amazon-titan sibling.

The meta-point is the same one that holds across Bedrock: every image model sits behind the same account, IAM, VPC, billing, and credit setup, so you can enable several, A/B them on your own prompts and brand, and route per use case — cheap Stability Core or Nova for bulk drafts, Stable Image Ultra for hero photorealism, Nova Canvas where the built-in watermark and editing suite fit — without re-plumbing anything. Benchmark on your imagery rather than on sample galleries, since relative quality shifts with each model generation.

VIIUse cases — marketing, product, and design

Image generation on Bedrock shows up most in three places: marketing content, product imagery, and design workflows. Mapping each to the right Stability tier keeps quality high and per-image cost sane.

• Marketing — ad creative, social, and campaign visuals: Generate on-brand ad variations, social posts, blog headers, and landing-page hero images at the volume modern testing demands. Draft broadly on Core (cheap, fast, many candidates), then re-render the winners on SD3.x or Ultra for the assets that ship. SD3.x's improved typography helps for visuals that need legible in-image text. Image-to-image keeps a campaign visually consistent from a reference.
• Product — catalog, lifestyle, and synthetic imagery: Produce product-in-context lifestyle shots, background variations for an existing product photo (image-to-image / inpainting to swap the scene while keeping the product), and large catalogs of consistent imagery. Inpainting and outpainting handle background replacement and reframing for different placements without a reshoot. Ultra is worth its price for hero product shots where photorealism sells.
• Design — concepting, mood boards, and assets: Speed up early-stage design: concept exploration, mood boards, texture and pattern generation, icon and illustration drafts, and rapid visual iteration on a direction. The SDXL-era model's style presets and fine control suit designers who want to steer the look; pinned seeds make controlled variation repeatable. Cheap per-image cost makes wide exploration affordable.
• Tiered generation — draft cheap, finish premium: The highest-leverage pattern, the image analogue of tiered text routing: generate a wide batch of candidates on the cheapest model (Core or the SDXL-era model), select the strongest compositions, then re-render only those on Ultra. Because cost is strictly per image, concentrating the expensive renders on the few finalists is where most of the savings live.

VIIILicensing, commercial use, and content notes

Generated imagery raises questions text does not: can you use it commercially, who owns it, and what guardrails apply. Here is the honest, practical orientation — none of this is legal advice, and you should confirm current terms before you ship.

Commercial use. Images generated through Stability models on Bedrock are generally usable commercially under the AWS service terms and the model provider's terms as surfaced on Bedrock — which is a key reason teams generate via Bedrock rather than self-hosting a model under a more restrictive standalone licence. That said, licensing terms for AI image models have shifted repeatedly across the industry, so before you build generated imagery into a paid product or paid campaign, read the current Stability model terms and the AWS service terms for the specific model, in your region, as of when you ship. Treat the per-model terms in the Bedrock console / AWS legal terms as the source of truth.

Ownership and IP risk. The legal status of who owns an AI-generated image, and the copyright treatment of model outputs, remains unsettled and varies by jurisdiction — in some jurisdictions purely AI-generated images may not attract copyright protection at all. For brand-critical or trademark-sensitive assets, many teams use generation for ideation and drafts and keep a human in the loop for final, published work. Avoid prompting for living artists' names, recognizable trademarks, or real people's likenesses in commercial output.

Watermarking and provenance. Provenance is an active area: Amazon's own image models (Nova Canvas, Titan Image Generator) embed an invisible watermark on every generated image by design, and the industry is moving toward content-credential standards (such as C2PA) for AI-generated media. If disclosure or provenance of AI-generated imagery matters for your use (regulated advertising, news, platforms with AI-content policies), factor watermarking and content-credential support into your model choice and pipeline.

Content safety. Bedrock's image models apply content filtering, and you can layer additional review into the pipeline for brand-safety and policy compliance. For user-facing generation (letting your own customers create images), plan for moderation of both prompts and outputs — the same governance discipline you would apply to any user-generated content, now with the IAM, logging, and audit trail that running on Bedrock gives you.

IXWho pays for it: partner-filed funding and Activate credits

Everything above prices Stability on Bedrock if you pay AWS directly. Two kinds of AWS funding can offset it: funding a partner files for you when the work qualifies, and Activate credits you apply for yourself.

Partner-filed funding. If the work qualifies, your matched AWS partner can file POC (proof of concept) funding to test an image-generation pipeline, or incremental workload (IW) funding if you already run on AWS and are launching new work. If you are moving to AWS from another cloud, it can file MAP migration funding. For startups this is typically $50K to $150K. Depends on the workload; AWS approves. Only a partner can request it, and it is tied to real work the partner does with you. See AWS partner funding.

AWS Activate credits. You apply for these yourself at aws.amazon.com/startups/credits: Activate Founders gives $1,000 to start, up to $5,000; Activate Portfolio gives up to $200,000 with an Org ID from your VC, accelerator or angel group if it is an Activate Provider. A consulting partner cannot file Activate for you. Bedrock usage generally draws down promotional credits, but check your credit terms: AWS promotional credits do not pay Savings Plans or Reserved Instance upfront fees.

CloudRoute matches you with one vetted AWS partner that can file that funding and help build the image workload: a generation API behind a queue, draft-on-Core and finish-on-Ultra routing, image-to-image and inpainting, storage and delivery, and moderation. Startups pay CloudRoute nothing; partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding. If you only need Activate credits, apply yourself; you do not need us.