Midjourney vs. Stable Diffusion for creator control

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of Midjourney or Stable Diffusion. What follows rests on official documentation, published pricing, licences and court records we read in full, all accessed on 16 October 2026. A tested update will follow.

Midjourney and Stable Diffusion are usually compared on image quality, which needs a tested review. The more durable difference is who controls what.

Midjourney is a hosted service: you rent a closed model, steer it with the tools Midjourney offers, and accept its terms on visibility, moderation and change. Stable Diffusion is a family of open-weight models from Stability AI. You can call them through Stability's API, or run them on your own hardware, choosing the tools, the version and where your files go, within a licence that still has conditions.

This comparison works through controls, custody, licences, privacy, cost and litigation. For the wider field, see our shortlist of AI image generators for creator workflows.

What each is for

Midjourney describes its versions as different models released over time, and its current default is V8.2, released on 24 July 2026.[1] You work on midjourney.com or in Discord, and the terms forbid automated access, so there is no sanctioned way to script it.[2]

Stable Diffusion is the model family; Stability AI publishes the weights. Its "Core Models" page, last updated 20 May 2026, lists Stable Diffusion 3.5 Large, Large Turbo and Medium as the newest image models.[3] Stability's own image page still headlines 3.5 as "Our most powerful image model yet".[4] Older releases such as SDXL carry a different licence (CreativeML Open RAIL++-M).[5]

Sources: Midjourney and Stability AI documentation, licences and pricing cited below.[1][6][2][7][8][9][3][10][11][12] US list prices, excluding tax.

Midjourney's controls: references, edits and video

Midjourney's controls are extensive, but each is one it designed, priced and can change.

Style references apply "the overall style—like colors, medium, textures, or lighting" of an image without copying objects or people.[13] Weight runs from --sw 0 to 1000, and numeric style codes from Midjourney's library can be reused or mixed.[13]

Character and object references have moved twice. Character Reference (--cref) is now a legacy V6 feature; Omni Reference (--oref), which put "characters, objects, vehicles, or non-human creatures" from one image into a scene, is V7-only and costs twice the GPU time.[14][15] In V8.1 and V8.2 both are replaced by the Edit model, which edits images from written instructions, composes with up to four reference images, and powers inpainting and outpainting.[16][1] Midjourney warns that fine details "like specific freckles or logos on clothing may not perfectly match your reference".[15] Our guide to keeping characters consistent in AI video covers why reference images reduce drift rather than remove it.

The Editor offers erase-and-regenerate brushes, Smart Select masks, layers, canvas expansion and a Retexture-style restyle, for Midjourney images or your own uploads.[17] One detail matters for privacy: Editor creations made with the V8 Edit model appear publicly on midjourney.com "unless you have Stealth mode enabled", and "This applies to Midjourney-generated images and external images."[17]

Video turns a start image into a five-second clip, extendable four seconds at a time up to 21 seconds, with optional end frames, loops and low or high motion.[18] Output is 480p, or 720p on Standard and above; a batch of four HD clips costs 26 GPU minutes.[18]

There is no documented way to train Midjourney on your own images, beyond Personalization profiles and Moodboards built from your ratings and picks.[1]

Stable Diffusion's controls: the open pipeline

With the weights on your machine, control comes from an ecosystem rather than a menu. Three pieces do most of the work.

ControlNet adds "spatial conditioning controls" to a pretrained model, guiding output with edges, depth, segmentation or human pose, according to the 2023 paper that introduced it.[19] Stability released Blur, Canny and Depth ControlNets for SD 3.5 Large under the same Community Licence.[20]

LoRA, low-rank adaptation, freezes a model's weights and trains small added matrices instead, cutting the cost of fine-tuning.[21] The technique was introduced for language models; Stability's licence names "low-rank adaptation" models as derivative works, which licensees own.[11]

ComfyUI, the open-source node-graph interface Stability's model card recommends for local use, chains models, LoRAs, ControlNets, masks and upscalers into reusable workflows.[10][22] It saves the full workflow and seeds inside generated images, runs on NVIDIA, AMD, Intel and Apple Silicon hardware, and its README says it "Runs fully offline".[22] It also supports many non-Stability open models.

The trade is effort: you assemble and maintain this stack. If you would rather not, Stability's API sells fixed "Control" endpoints (Structure, Sketch, Style) at 5 credits a call, which buys convenience but returns custody to the vendor.[9]

AI Chronicle

Who holds the model: hosted service vs. open weights

The figure shows the pattern. With Midjourney, the vendor holds the weights, the servers and the roadmap. Its terms say the service may change, "including but not limited to the art style of Assets", and ask users not to "create any dependencies on any attributes of the Services".[2] Versions do retire: V8.0 "is no longer available for use".[1] Midjourney can also "suspend or ban Your access to the Services at any time, and for any reason."[2]

Self-hosting reverses most of that. A downloaded checkpoint does not change unless you replace it, and a workflow saved with its seed can be rerun later. Open weights do not mean no rules, though. As our explainer on open versus closed AI models sets out, the licence still decides what you may do. The Stability AI Community Licence binds outputs to its acceptable use policy, forbids using outputs "to create or improve any foundational generative AI model", and is "revocable" if you breach it.[11][10]

Licences and commercial terms

Midjourney. You "own all Assets You create with the Services to the fullest extent possible under applicable law", with exceptions: a company, or an employee of one, with more than $1 million a year in revenue must be on Pro or Mega to own its assets, and upscaled images of other people's work remain theirs.[2] In return you grant Midjourney "a perpetual, worldwide, non-exclusive, sublicensable no-charge, royalty-free, irrevocable copyright license" over your inputs and outputs.[2] Disputes go to binding arbitration, and you indemnify Midjourney against third-party claims arising from your use.[2]

Stable Diffusion. The Community Licence allows research, non-commercial and commercial use free of charge, but commercial users must register.[11] The revenue test is wider than it looks:

“regardless of whether that revenue is generated directly or indirectly from the Stability AI Materials”

— Stability AI Community License Agreement, Section III (last updated 5 July 2024).[11]

Revenue is counted with affiliates. Cross $1 million and the licence terminates; an Enterprise licence is granted at Stability's "sole discretion".[11] Distributors must display "Powered by Stability AI".[11] You own your fine-tunes and, as between you and Stability, your outputs; the licence terminates if you sue Stability for infringement over them.[11]

Neither vendor guarantees outputs are non-infringing; both provide them "as is".[2][11] Whether anyone can own a purely prompted image is a separate question, covered in our guide to AI for marketing.

Privacy and visibility

"By default, Your Content is publicly viewable and remixable," Midjourney's terms say.[2] Stealth mode, on Pro and Mega only, commits Midjourney to "best efforts" not to publish your assets, but anything made in a public Discord channel stays visible.[2][7] Stealth does not retroactively hide earlier public work.[7] Its privacy policy lists prompts and uploaded images among the personal data it collects, and notes ad-cookie sharing you can opt out of.[8] We found no documented opt-out from training.

Stability's hosted services say they "may use Content to improve and develop our Services", with an opt-out from training on your inputs and outputs.[23][24] Self-hosting removes the vendor from the loop entirely, which matters for unreleased client work. Provenance is a separate step: our explainer on content credentials covers how to label what you publish.

Cost model and hardware

Midjourney charges a subscription measured in GPU time: 3.3 Fast hours on Basic up to 60 on Mega, extra time at $4 an hour, and unlimited Relax-mode images from Standard.[6] Video in Relax mode needs Pro or Mega.[6]

Self-hosting swaps the subscription for hardware. Stability says SD 3.5 Medium needs 9.9 GB of VRAM, "excluding text encoders", for full performance. SD 3.5 Large, at 8.1 billion parameters, ships with guidance on 4-bit quantisation and CPU offload for smaller GPUs.[10][12] ComfyUI claims it can run large models "on as low as 4GB vram + 8GB ram" using weight streaming.[22] These are vendor claims, not measurements. Without a suitable GPU, Stability's API is the pay-per-image middle path: 6.5 credits ($0.065) per SD 3.5 Large image, 3.5 for Medium, with 25 free credits to start.[9]

Where the litigation stands

None of these cases has produced a verdict on whether training these models is lawful. Our guide to AI training and copyright explains the legal questions.

Getty Images v Stability AI (UK). Getty dropped its training and output claims before judgment. On 4 November 2025, Mrs Justice Joanna Smith rejected its remaining copyright claim, holding that Stable Diffusion "does not store or reproduce any Copyright Works (and has never done so)".[25] Getty won only on trade marks, in findings "both historic and extremely limited in scope".[25] Getty's August 2026 quarterly report says its appeal on secondary infringement "is expected to be held in November 2026", and that the Court of Appeal refused Stability permission to appeal the trade-mark ruling.[26]

Andersen v. Stability AI (US). Filed in January 2023 against Stability, Midjourney, DeviantArt and Runway, this artists' class action is still in discovery. On 28 September 2026 Judge Orrick moved class certification and summary judgment to a hearing on 29 September 2027, left the trial date to be set, and wrote: "I do not expect to grant any further continuances."[27][28]

Disney and Universal v. Midjourney (US). Filed in June 2025 and consolidated with a Warner Bros. Discovery suit, this case alleges infringement through both training and user outputs.[29][30] Midjourney says it will establish that its use is "a quintessential 'fair use'".[30] Its motion for partial judgment on the pleadings against the contributory-infringement claim, citing the Supreme Court's March 2026 Cox decision, is set for 23 November 2026; a training-data production order was approved on 29 September 2026.[29][30] There is no ruling on the merits.

Documented limits and risks

• Midjourney can change under you. Features have been replaced twice in two versions, and the terms disclaim any guarantee of stability.[1][2]
• Public by default. Lower plans cannot hide work, including V8 Editor edits of your own uploads.[17][7]
• Moderation friction. Midjourney warns that "seemingly innocent prompts may be blocked".[15]
• The $1 million test bites differently. Midjourney ties it to ownership of outputs; Stability ties it to the licence itself, counting all revenue and affiliates.[2][11]
• Self-hosting shifts safety to you. Stability's model card tells developers to "implement content safety guardrails" for their own use cases.[10]
• Stability's model roadmap is unclear. We found reports of a newer "Stable Diffusion 4", but no such model on Stability's own pages.[3][4]

Who it suits, and who should skip it

• Midjourney suits illustrators, art directors and marketers who want reference-driven iteration and video without managing hardware. Skip it if client work must stay private below Pro, if you need scripted pipelines, or if the model must not change mid-project.
• Self-hosted Stable Diffusion suits studios and technical artists who need exact composition, custom LoRAs, offline confidentiality and reproducible workflows. Skip it if you lack a capable GPU or time to maintain a node stack, or are near $1 million in revenue without an Enterprise budget.
• Stability's API suits developers who want Stable Diffusion without running GPUs.

How to test it yourself

1. Write a five-image brief that needs a fixed character, a fixed layout and a house style.
2. Check visibility first: in Midjourney, confirm your plan and Stealth setting before uploading anything confidential.
3. Run the brief with references in Midjourney (style reference plus Edit-model references) and record how many generations each image took.
4. Run it locally with SD 3.5 and a pose or depth ControlNet, saving the ComfyUI workflow and seeds.
5. Rerun both a week later with the same inputs; note what changed.
6. Log the full cost, including setup time.
7. Check the licence against your revenue, counting affiliates.

A worked example (hypothetical)

This scenario is invented to show how the choice works. No tools were run and no outputs are reported.

Amna runs a three-person studio making a picture book and its launch campaign. The studio's revenue is well under $1 million.

Concept art goes to a hosted service. For style exploration they use style references and the Edit model on a Pro plan, with Stealth on, because the characters are unannounced.

Final pages run locally. Once the character is fixed, they train a LoRA on their own approved drawings and use a pose ControlNet per spread, in a saved ComfyUI workflow. A change requested in six months can reuse the same model, seed and graph.

Paperwork is set once. They register for Stability's Community Licence, add the required attribution to the campaign site, keep records of human edits, and review whether any output resembles a known character.

Their test for each tool: does it give control where the project needs it, and custody where the client needs it?

Questions readers commonly ask

Is Stable Diffusion free?

The weights are free to download. Commercial use is free below $1 million in annual revenue, counted across your affiliates, after registering; above that you need an Enterprise licence.[11][10] Your hardware or API credits are extra.

Can I keep Midjourney images private?

Only on Pro or Mega, through Stealth mode, and not in public Discord channels.[7] On other plans, work is publicly viewable and remixable by default.[2]

Can Midjourney be trained on my own style?

Not with custom weights. It offers style references, style codes, Personalization and Moodboards.[13][1] Training your own LoRA is a Stable Diffusion workflow.

Can I use Stable Diffusion without a powerful GPU?

Yes, through Stability's API or cloud partners, at the cost of sending prompts to a vendor.[9][4] Locally, ComfyUI says it can stream weights to fit small cards; expect slower generation.[22]

The next step

Both companies are defending lawsuits over how their models were trained. Next, read AI training and copyright: the important distinctions to see which legal questions those cases can and cannot settle.

Sources and evidence note

Product features, defaults, prices and terms come from Midjourney's and Stability AI's own documentation, licences and pricing pages, and describe intended behaviour, not tested performance. Hardware figures are vendor or project claims. Litigation status comes from court judgments, dockets, filings and a company securities filing. Prices are US list prices as displayed on 16 October 2026. The worked example is invented. All pages were accessed on 16 October 2026.

Related reading

• Best AI image generators for creator workflows
• How to spot AI-generated content: signals and limits