One model family sold on a look, with V8.2 current and video running inside the same plans.
Midjourney sells a single model family and nothing else. V8.2 is current, reachable on the web and on Discord, and video generation is included rather than sold separately. What a plan really buys is throughput: three fast image jobs at a time on the cheaper two tiers, twelve on the pricier two. Privacy follows the same ladder, since below the top two tiers the community can see what you make.
Best for: A signature look across a series
- One subscription covers the image models and the video ones.
- Stealth mode, which hides your generations from other users, starts on Pro.
- Nothing is free to try; the entry point is a paid plan.
- A company over $1M gross a year has to hold one of the top two plans.
An open image model now aimed at products and agents, still known for setting readable type.
Ideogram sells itself to products and agents these days, though type inside the frame is what its name still carries. Past generation, the capability set runs to upscaling, transparent layers, editing controls, prompt fidelity, and a custom model you train yourself, with an MCP surface beside the web app and the API. The model runs on Morphic as well, so picking it does not mean a second subscription.
Best for: Type that has to be legible
- Three surfaces reach the model: the web app, the API, and an MCP server.
- Output is cleared for any use, with no revenue threshold attached.
- Ten credits a week on the free plan, and they run at slow priority.
- The product stops at images; anything that moves or sounds needs another tool.
Adobe's generative workspace, sold on a legal guarantee that covers part of what sits in the picker.
Firefly puts Adobe's own image, video, and speech models in the same picker as models from outside labs, and reaches into Photoshop, Premiere, and the rest of Creative Cloud. The IP indemnity everyone cites is enumerated in a legal product description that excludes anything marked beta or powered by a non-Adobe model. On a paid plan standard image work deducts nothing, while video, audio, avatars, and the partner models are metered.
Best for: Work that has to clear legal
- The indemnity is a contractual commitment, not a marketing position.
- C2PA Content Credentials are applied automatically at export.
- Pictures made with the partner models fall outside the indemnity.
- A credit balance resets each month, and every person keeps a separate one.
The checkpoint is yours to download, and the license costs nothing until revenue crosses a line.
Stable Diffusion 3.5 remains Stability AI's flagship image family: a Large build at 8.1 billion parameters, a four-step Large Turbo distill, and a 2.5 billion Medium that Stability puts at 9.9 GB of video memory, excluding text encoders. The Community License charges nothing for commercial work below one million dollars of yearly revenue. Hosted first-party access means the API, Stable Assistant, or Brand Studio, since DreamStudio has closed.
Best for: Keeping the model on your disk
- The weights run locally, so no vendor can withdraw a checkpoint you hold.
- Fine-tunes, LoRAs, and retrains are expressly permitted, including selling those derivatives.
- No ongoing free first-party tier; the API grants 25 credits and the apps run trials.
- The license requires visible attribution and can be revoked on breach.
A hosted toolkit where each stage of making a picture has its own named module.
Leonardo AI gives every stage its own tool. A mask painted over the wrong part of a picture sends only that region back for another attempt, sketching drives a live preview so composition is settled by drawing, and Flow State scrolls variations for the open-ended part of a job. Recurring characters come from Personal AI Models, which are LoRA fine-tunes, and from Blueprints.
Best for: Deciding a picture by drawing it
- The live preview turns a rough sketch into an image as you draw it.
- Personal AI Models and Blueprints give a repeatable route to character consistency.
- Editing is mask-based: you paint the area to regenerate rather than isolate an element.
- Audio arrives inside the video output rather than as tracks to arrange.
A canvas that answers while you are still typing, plus the upscalers that finish the job.
Krea makes generation feel continuous. The picture redraws under the cursor as the prompt or the sketch changes, which is a different way of working from submitting a job and waiting on it. Its own models sit alongside third-party ones, and its upscaling is good enough that people keep an account purely for finishing pictures made elsewhere. Krea models are on Morphic's roster as well.
Best for: Working out a direction quickly
- Changes register as they are made, so a direction gets ruled out in seconds.
- Upscaling and enhancement handle finishing work on pictures generated anywhere.
- The canvas is tuned for single frames, not for assembling a sequence.
- The roster mixes Krea and third-party models, so results vary with the pick.
A design-side image tool where raster and vector come out of the same generation.
Recraft is aimed at designers more than prompt artists. Its current model line returns raster or vector from one app, typography holds up on poster and logo work, and a team can train a private brand style on its own assets so a series stays consistent. For brand and product imagery after DALL-E, it is the most design-shaped landing spot on this list.
Best for: Logos, posters, and vector files
- Vector export covers the formats a brand system actually ships.
- A private style trained on your assets keeps a campaign consistent.
- Art-direction range is narrower than the general-purpose models.
- There is no video path, so motion work needs a second tool.
