Four jobs, not forty tools
Image generation looks like an endless list of products. In practice, teams hire these tools for four distinct jobs, and almost every subscription in the category is a duplicate of one of them. Sorting your spend by job is the fastest way to find the tool you are paying for twice.
Job one is where the money goes and where the least risk lives, because switching between two hosted generators costs you a weekend of prompt rewriting. Job three is the opposite: the model cannot be revoked, but you inherit GPU cost, upgrades and safety filtering. Job four is the quiet trap - generation inside a design editor is convenient right up to the day you need the layered original outside it.
The status picture for this category
Of the eleven tools tracked in this collection, seven are actively maintained, one was absorbed into a larger company, one is abandoned, one has fully shut down, and one is showing enough neglect that we class it as possibly dead. That is a roughly one-in-three failure rate for a category that is supposedly the healthiest in consumer AI.
The failures cluster in one place: products built on a single viral moment. Lensa AI had the avatar summer, DALL-E Mini had the meme era, and Disco Diffusion owned the 2022 art scene before hosted competitors made its notebook workflow obsolete. None of them failed on quality. They failed because a spike in demand is not a retention curve.
Rights, indemnity and the paperwork nobody reads
The single biggest difference between these tools is not image quality - it is who carries the legal risk. Adobe built Firefly on licensed and public-domain material specifically so enterprise buyers could get indemnification, and publishes its commercial-use terms in plain language. Most competitors put the risk on you.
| Tool | Best at | Commercial clarity | Lock-in risk |
|---|---|---|---|
| Midjourney | Aesthetic quality, style references | Paid tiers grant broad rights; no indemnity | Medium - gallery and seeds live server-side |
| DALL-E 3 | Literal prompt adherence, text in images | Ownership granted; policy-driven refusals | Low - available through the OpenAI API |
| Stable Diffusion | Fine-tuning, offline and batch work | Open weights; licence varies by checkpoint | Very low - you hold the model |
| Adobe Firefly | Indemnified enterprise output | Strongest - contractual indemnification | Medium - tied to Creative Cloud |
| Stability AI | Open-model access and research | Model-by-model, read each licence | Low on weights, higher on hosted APIs |
| Canva Magic | Generating inside an existing design | Tied to Canva plan terms | High - assets and layers live in the editor |
| Jasper Art | Marketing-adjacent visuals | Bundled into Jasper plans | High - see the postmortem below |
If you are producing anything that will appear in paid media, indemnity is not a nice-to-have. If you are producing concept art, mood boards or internal decks, it is largely irrelevant and you should optimise for quality instead. Deciding which of those two you are before you compare tools removes about half the shortlist.
The death signals that actually predict a shutdown
We score every tracked tool against the same signals. In image generation, they carry different weights than in other categories, because a stale checkpoint is a much louder alarm here than a quiet changelog would be for a database product.
Artbreeder is the clearest live example: the product still loads and accounts still work, but the generative core has not meaningfully advanced while competitors shipped several model generations. That is the profile of a tool that will keep taking your money for another year and then close with thirty days notice. Users have been drifting to the options in tools like Artbreeder.
What actually happened to the dead ones
Four cautionary tales, each with a different failure mode:
- Disco Diffusion - a Colab notebook that defined the 2022 aesthetic and lost to one-click hosted products. Full story in is Disco Diffusion dead and replacements in tools like Disco Diffusion.
- DALL-E Mini - the meme engine whose team was absorbed, leaving the public demo behind. See is DALL-E Mini dead and tools like DALL-E Mini.
- Lensa AI - the fastest rise and quietest stall in the category, a single viral avatar feature with nothing behind it. Status on the Lensa AI page.
- Jasper Art - a generator bolted onto a copywriting suite, then de-emphasised when the parent refocused. The postmortem is in why Jasper Art failed.
Two of these were never businesses; they were features with a growth spike. The other two were bundles, and bundled features are always the first thing cut when the parent company reprioritises. That pattern generalises well beyond images - it is the same story we documented for Stability AI, where the models outlived the company''s original commercial plan.
A 30-day audit for your image stack
Run this once and you will not be surprised by a shutdown notice again.
- Days 1-3: inventory. List every image tool with a live subscription, the card it bills to, and the one job it does. Anything without a job is cancelled now.
- Days 4-7: export. Pull full-resolution originals from every tool, including upscales and variants. Confirm the archive opens outside the vendor.
- Days 8-12: capture the recipe. For your top fifty images, record prompt, seed, model version and reference inputs. Without these, a backup is decoration.
- Days 13-16: check the paperwork. Confirm your plan grants commercial rights for how you actually use the output, and note which vendors offer indemnity.
- Days 17-21: test the replacement. Reproduce five signature images on a second tool. Measure how close you get, not whether it is possible.
- Days 22-26: cut duplicates. One tool per job from the map above. Downgrade rather than cancel where you still need archive access.
- Days 27-30: set the alarm. Follow each remaining tool on its status page here so a change in verified status reaches you before an email from the vendor does.
Which tool for which job
The honest comparison is not a feature list, it is a strength profile. Every tool in this category is strong at something and mediocre at something else, and the mediocre part is where teams get stuck six months in.
Read it as a shortlist generator. Need the best-looking frame and can live without indemnity? Compare Midjourney against DALL-E 3 and pick on style. Need paperwork you can hand to a client? Firefly, and accept the aesthetic ceiling. Need to be immune to shutdowns? Self-host, and accept the ops work.
Where the category is heading
Two forces will decide the next round of failures. The first is model commoditisation: open checkpoints now match hosted quality for most briefs, which strips pricing power from anyone whose only asset was a wrapper. The second is legal, and the industry is watching cases and guidance such as the US Copyright Office AI reports and the transparency obligations in the EU AI Act. Vendors who cannot document their training data will lose enterprise accounts regardless of output quality.
Practically, that means the safe positions are the two extremes: a very large platform with legal cover, or a model you run yourself. The dangerous middle is a mid-sized hosted product with a licensed-from-somewhere model and no distribution - which is exactly the profile of every tool in the graveyard section above.
If you want to keep auditing beyond images, the same method applies to adjacent stacks: see AI tools for marketers, AI tools for writing blog posts, or browse the full collection index. Every status on this site is re-verified monthly, and you can report a change if you spot something before we do.