Hosted platform where you can still train a personal model and edit at the level of a mask.
Leonardo runs its own Lucid Origin and Lucid Realism models beside third-party ones including Flux, Ideogram, Seedream, and Nano Banana Pro, covering image, video, editing, and upscaling in one app. Personal model training takes a small set of images, with 10 to 50 trained models a month depending on plan.
Best for: Training your own hosted model
- Personal model training keeps the fine-tuning habit alive without a local GPU
- Unused tokens roll into a bank rather than expiring at the billing reset
- Free-plan creations are public, and Leonardo holds rights to use and distribute them
- Simultaneous generations are capped by tier at two, three, or six
#3
FLUX (Black Forest Labs)
The open-weights image line that is still shipping checkpoints you can download and fine-tune.
Black Forest Labs sells FLUX.2 through an API and publishes FLUX.2 klein for local use in four builds. The 9B distilled model and the 9B base carry the FLUX Non-Commercial License; the 4B pair is Apache 2.0. Weights sit on Hugging Face, the inference code on GitHub, and a signup-free demo runs in the browser. Flux models are on Morphic's image roster as well.
Best for: Open weights you can still download
- Two klein builds are Apache 2.0, with no revenue threshold and no attribution condition
- The smallest build asks for 8.4 GB of VRAM and the largest for 21.7 GB
- The 9B builds are non-commercial, so paid work runs on the 4B builds or the API
- No first-party creative app, so hosted use goes through the API or another platform
Multi-model canvas that redraws the picture in real time as the prompt or the sketch changes.
Krea turns generating into something closer to drawing: the canvas regenerates while you type or sketch, so a direction is judged in the moment instead of one submitted job at a time. Its in-house models sit in the same picker as ones it licenses, and its upscaling is strong enough to be the finishing pass. Krea models also appear on Morphic's image roster.
Best for: Judging a direction in real time
- Real-time regeneration replaces the submit-and-wait loop with something continuous
- Upscaling and enhancement finish an image rather than only start one
- Scope ends at generating and finishing a picture, with no multi-clip timeline behind it
- The free tier carries no commercial license, and video generation there waits on paid concurrency
Image model whose maker now leads on openness rather than on the poster typography that made its name.
Ideogram 4.0 is sold by its maker as an open model built for visual intelligence, a turn away from the text-rendering app most people met first. Its listed capabilities run from prompt fidelity through editing controls and upscaling to layers with transparency and models you create yourself. Ideogram is on Morphic's image roster.
Best for: Open-model access with edit tooling
- Every output may be used commercially, with no revenue threshold attached
- Models you create yourself and layers with transparency live in the same product
- Ten slow credits a week is what the free plan hands out
- Nothing exists past the picture: no video, no audio, no assembly surface
Adobe's creative surface, where its own models carry a contractual IP indemnity and partner models do not.
Firefly puts Adobe's image, video, and speech models beside a partner roster reaching across Google, OpenAI, Kling, Runway, Luma, and Black Forest Labs. The indemnity Adobe is known for is enumerated in its legal product description, which expressly leaves out anything powered by a non-Adobe model or marked beta. Paid plans run from US$9.99 to US$199.99 a month.
Best for: Client work that needs an indemnity
- Generative Fill and the rest of the standard image set cost nothing on a paid plan
- Content Credentials attach at export, which procurement processes increasingly ask for
- Whatever is left over is gone at the reset, and each member draws from a personal allocation
- Video is metered per second, so 1080p costs 100 credits for every second generated
The aesthetic reference point for AI images, on a subscription with no free way in.
Midjourney runs V8.2 as of August 2026 across a web app and Discord, and its video generation is live rather than promised. Concurrency follows the plan: three fast image jobs at once on the entry and mid tiers, twelve on the two above them, with the waiting line stopping at ten. Stealth mode, its name for private generation, begins on Pro, and companies over $1M in gross revenue must sit on Pro or Mega.
Best for: Aesthetic-led art and concept work
- Relax mode allows unlimited image generation from the Standard plan upward
- Video runs on the same plans as images, with no separate product to buy
- There is no free tier, and every plan starts as a paid subscription
- Private generation begins on Pro, so work below that tier is visible to the community
#8
DALL·E (ChatGPT Images)
OpenAI's image line, with the DALL·E name retiring on 30 August 2026 and generation continuing as ChatGPT Images.
The DALL·E GPT is scheduled to leave ChatGPT on 30 August 2026, a date OpenAI put in writing on the last day of July, and anyone still generating there is sent to ChatGPT Images instead. That surface runs the GPT Image family, gpt-image-2 being current, and both older API models were withdrawn on 12 May 2026. Editing survives the change: highlight an area in the chat and describe the fix, or call the Images API.
Best for: Image generation inside a chat assistant
- A picture on the zero-dollar tier carries no mark you can see, and the terms hand ownership to whoever generated it
- The developer surface is documented and covers generation, edits, and batched jobs
- The DALL·E name goes away on 30 August 2026, so anything built around it moves to ChatGPT Images
- Training on your content is the default on consumer plans, with an opt-out; ChatGPT Business starts from the opposite position
