the idea
the independent-lab bet — the lineage midjourney inherited, what it proved about taste and distribution, and what the world-model ambition requires.
lineage
- 1972–2010saaronharold cohen's forty-year program — the first serious ai art system, authored by an established painter asking 'what are the minimum conditions under which a set of marks functions as an image?' the whitney's 2024 retrospective explicitly positioned it as the ancestor of 'dall-e, midjourney, and stable diffusion.' the lineage break: cohen's machine collaborated with its artist; midjourney's 'collaborators' never consented — the gap that fuels the litigation.
- 2015deepdreammordvintsev's inception-derived visualizer went viral june–july 2015 and open-sourced july 1 — the first mass-audience proof that a trained classifier run in reverse produces the machine's own aesthetic, and the template for the ai-image virality midjourney later industrialized.
- 2014–2018gan art and the belamy salegoodfellow's gans → robbie barrat's portrait code → obvious's 'edmond de belamy' selling at christie's for $432,500 in october 2018. the auction is the direct ancestor of every later ai-art controversy: unattributed borrowed code, press-release marketing, an auction house validating 'the algorithm' as author — the authorship/credit fight midjourney reignited at scale.
- 2021–2022the diffusion + clip wavelatent diffusion (rombach/blattmann/esser/ommer, cvpr '22) + laion-5b + clip guidance collapsed the cost of text-to-image; dall·e 2 stayed gated, imagen stayed unreleased, stable diffusion went open-weights on 2022-08-22. midjourney — a ~10-person lab training diffusion models from september 2021 — rode the same wave but shipped the product: a discord bot, public channels, $10/mo. the science was communal; taste and distribution were the differentiators.
- ~2008–2021the spatial-computing prehistoryholz's pre-history — physics and math, a fluid-mechanics phd left mid-program for nasa/max planck research, then twelve years building leap motion's hand tracking: 'the right product at the wrong time.' he left in 2021 explicitly to escape venture-backed company-building. the name is zhuangzi's 中道 via a favorite translation — 'mid-journey.'
- 2021–the craigslist modelthe business-model ancestor — craigslist: a weird, standalone, profitable internet business that never raises and never explains. holz cited it directly for what midjourney should be: 'this weird thing that no one knows how to compete with.'
what midjourney proved
- the self-funded outliera self-funded ~40-person lab can out-earn venture-backed rivals — reported $200m+ revenue in 2023, profitable early, zero marketing spend, zero dilution; disney's own complaint alleged $300m for 2024. advisors (friedman, keller, rosedale) instead of a board. all figures reported or litigant-cited — but even discounted, the counterexample stands.
- taste as moattaste as moat — the opinionated default aesthetic ('reaches for beauty': flattering light, composition, mood) won preference tests and became the reference 'ai look.' personalization, moodboards, and sref industrialized taste as a feature; v7's forced 200-image taste profiling literalizes the moat. competitors benchmark against it even while beating it elsewhere.
- discord as distributiondiscord as distribution — the largest server the platform ever hosted (~21m members by late 2024, up from ~13m in march 2023); discord had to engineer past its own 1m-member cap. public-by-default generation was viral prompt education, free curation data, and word-of-mouth acquisition in one loop — a chat server as both product and growth engine.
- the community as flywheelthe community is the dataset flywheel — public prompts and outputs teach users, seed new aesthetics ('aesthetic accelerationism'), and produce the ranking signal personalization runs on. the community-facing design doubles as r&d.
- the small-lab velocityproduct velocity at ~60 people — v1 through v8.2, a video model, a web editor, rooms, patchwork, and a pytorch rewrite, shipped by a lab whose own site describes 'a lean, self-funded, distributed team.'
what stayed unfinished
- the text problemin-image text arrived late — ideogram solved legible text ~seven months before v6 shipped only 'minor text drawing ability'; real strength waited for v8.1.
- two years on someone else's railsa real web product — discord-only until august 2024, two and a half years in. 'escape from discord island' was a relief precisely because the dependency had become the story.
- the free tier that comes and goesa durable free tier — suspended march 2023 for 'extraordinary demand and trial abuse,' restored august 2024 at ~25 images, gone again. the top of the funnel never stabilized.
- late to videovideo arrived ~18 months behind sora's demo and after veo, runway, and kling — and every generation is still image-anchored; text-to-video remains absent.
- no apino public api — the recurring enterprise complaint in every comparison; no official statement of intent found.
- the unshipped world modelthe world-sim promises stay promises — the march 2024 office-hours roadmap (3d, video, real-time → 'open world sandbox') and the fall 'holodeck' reporting still had no public artifact as of september 2026; v8.2 is still an image model and the edit model is tooling, not world-sim.
the world-model question
- forthe images were never the product thesis — 'we don't make images because there is a thing called art… the medium ends in simulated worlds.' the subscription image business is the cash engine funding the world-model ambitions; v1 video was framed on arrival as 'a stepping stone' to real-time open-world simulation. judged on that axis, every release is progress.
- forthe bootstrap is the strategy — no investors means no exit pressure, no board, and a ten-year arc that can afford weird. craigslist permanence is the stated model; 'raising money is very different than having a product that people are willing to pay for.'
- againstcan a ~60-person self-funded lab litigating three studio suits actually build world models? the sims vision implies orders of magnitude more compute and team; the march 2026 interview concedes hardware may force the first vc raise. the ambition-to-structure gap is the open question.
- againstthe fair-use exposure is asymmetric — holz's 2022 candor is plaintiffs' exhibit; a loss could force licensing economics onto a self-funded p&l, while a win ratifies scrape-training for everyone. midjourney's closed model may make it more exposed to output-liability theories than open-weight defendants.
- nuancetaste as moat vs. converging open weights — traffic decay says the casual funnel is gone, and chat platforms now own the viral function midjourney invented. the bet is that a paying professional base plus personalization lock-in suffices — a moat that shows up in churn data nobody outside the company can see.
successors
- ideogramthe toronto lab from ex-google brain/imagen researchers (norouzi, saharia, chan, ho) — solved legible text first, called 'a serious midjourney contender' on launch; v2 landed the week midjourney opened the web.
- black forest labs / fluxthe freiburg team that literally wrote stable diffusion — rombach, esser, blattmann — launched august 2024 with a $31m a16z seed, claimed sota over midjourney v6, and now powers grok's images. the open-weights mantle midjourney never wanted.
- recraftv3 topped the artificial analysis image arena above midjourney and openai; svg vector output opened a design-tool category midjourney doesn't contest.
- gpt-4o and imagen — the platformsthe viral function moved from discord servers to chat products: the ghibli moment showed openai and google absorbing mainstream image generation as a feature. midjourney is no longer the face of ai images — it remains the aesthetic reference point.
- cara — the counter-movementthe anti-scrape artist refuge founded by andersen plaintiff jingna zhang — the successor in the opposite direction: a social network built on the premise that the training corpus was never consented to. scraped three times in ten days in august 2026.
open questions
- does fair use hold? the disney/universal and warner bros. dockets are the highest-stakes test, and holz's own 2022 admissions are plaintiffs' exhibits. a loss could force licensing economics onto a self-funded p&l; a win ratifies scrape-training for everyone.
- can taste remain a moat when open weights converge on quality and chat platforms own distribution? the traffic data says the casual funnel is gone — the bet is paying professionals plus personalization lock-in.
- can a ~60–100-person self-funded lab build world models, 3d, and hardware while litigating three studio suits? the sims vision implies orders of magnitude more compute — and the 2026 interview concedes hardware may force the first vc raise.
- what is the equity and endgame? no priced round exists; the $10b figure was analyst arithmetic; the stated model is 'kind of like craigslist' — permanence as a weird standalone. acquisition and ipo optionality are deliberately nil.
- will opt-out regimes, licensing, or nightshade-class poisoning reprice the training corpus — and did midjourney's closedness make it more exposed to output-liability theories than open-weight defendants?
- is midjourney medical the world-model funding engine, a second company inside the first, or the business that finally forces outside capital — and what happened to 'orb,' the consumer-hardware track the scanner replaced?
- what happens to the ~21m-member discord server — asset or archive — as the product's center of gravity moves to the web?
- who owns what? no cap table, subsidiaries map, or niji/spellbrush terms have ever been disclosed — the corporate perimeter is a finding, not a footnote.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.