Published January 22, 2024 · By CampaignsLive · Insights
By the end of 2023, three image-generation tools had emerged as the dominant options for anyone trying to use AI imagery in commercial work. Stable Diffusion, Midjourney, and DALL·E (in its v3 release through ChatGPT) covered roughly ninety percent of the active use among professional creative teams. Each had distinct strengths, distinct weaknesses, and distinct working norms. This is a working retrospective on where each one actually fit in the brand creative production stack at the end of 2023, written from the vantage of having spent a year watching teams adopt all three.
Stable Diffusion: the workhorse with a learning curve
Stable Diffusion’s 2023 trajectory was, in retrospect, the most consequential of the three. The 1.5 release at the end of 2022 had established the tool as the open-source default; the SDXL release in mid-2023 brought it up to a quality bar that began to compete seriously with the commercial options. Compared to the others, Stable Diffusion offered three things they did not.
The first was control. Through ControlNet, LoRA fine-tuning, and a deepening ecosystem of compositional tools, Stable Diffusion let a team direct the model’s output in ways that were not possible in Midjourney or DALL·E. A team could pose a figure, hold a composition, transfer a style, or constrain a palette in a way that the closed tools did not allow.
The second was deployment flexibility. Stable Diffusion could run on a team’s own infrastructure, in a tenant-isolated environment, on a brand’s own GPUs. For brands with sensitive material — unreleased product imagery, talent under NDA, campaign concepts still in development — this mattered enormously. The closed tools required uploading every input to a third party. Stable Diffusion did not.
The third was extensibility. The open-source nature of the model meant a brand could fine-tune it on its own visual archive, build custom checkpoints, integrate it into existing pipelines. The level of tooling investment required was higher than the closed options; the ceiling was also higher.
The trade-off was operational complexity. Stable Diffusion in 2023 required a meaningful engineering investment to use well. Teams that did not have that investment either ran into the model’s defaults — which were noticeably worse than the closed options — or ended up reaching for Midjourney instead.
Midjourney: the highest output quality, the worst working interface
Midjourney in 2023 produced the most consistently striking single images of the three tools. Its v5 and v5.2 releases (and v6 by the end of the year) set the visual register that most “this looks like AI in a good way” content was using. For static imagery in the editorial, fashion, or atmospheric registers, Midjourney was the default reach.
The working interface was its main weakness. Midjourney ran inside Discord throughout most of 2023. The interface was iterative-by-design — generate, vary, upscale, vary again — and worked well for solo creatives or small teams but resisted integration into formal production pipelines. Asset management was effectively impossible. Versioning was manual. Collaboration was awkward.
Midjourney also could not be controlled in the same way Stable Diffusion could. There was no ControlNet equivalent until v6 brought partial parity. Prompting was an art rather than a discipline. The same prompt could produce strikingly different results across sessions. For repeatable, brand-consistent output, this was a problem.
The pattern that emerged among teams using Midjourney professionally was that it excelled at concept exploration — generating a hundred directions in an afternoon — and faltered at finished work. Teams used it as a moodboard generator and then translated the resulting direction into Stable Diffusion or a traditional production process for the actual asset.
DALL·E: the broadest reach, the lowest ceiling
DALL·E’s 2023 was defined by its integration with ChatGPT. The v3 release in October, accessible inside the chat interface, brought image generation to a wider user base than the other two tools combined. By the end of the year, a vast number of people had used DALL·E without knowing it specifically — they had asked ChatGPT for an image and an image had appeared.
For brand creative work, DALL·E v3’s strengths were in two specific areas. The first was prompt adherence. DALL·E v3 was, at the time, the best of the three at following a detailed prompt without veering into the model’s defaults. If a prompt asked for a specific composition with specific elements in specific arrangements, DALL·E v3 was more likely to deliver that than Midjourney or Stable Diffusion. The second was typography. Though far from reliable, DALL·E v3 could produce legible text within an image more often than the other two, which was useful for concept work that required headline integration.
The weaknesses were the inverse. DALL·E offered no control beyond the prompt — no compositional steering, no style transfer, no fine-tuning. Output quality at the high end was generally below Midjourney’s. The tool was unsuited to brand-asset-anchored work because Microsoft and OpenAI’s content policies were aggressive on logo and trademark adherence, sometimes blocking generation of work that involved brand elements the user had legitimate rights to use.
The pattern among brand teams was that DALL·E was the access tool — used for quick concept tests, for non-designer stakeholders to visualize an idea, for working sessions where the goal was alignment rather than asset production. It rarely produced final creative.
What the tools shared, and what they did not
All three tools, by the end of 2023, shared the same fundamental limitations. None produced reliable text within images. None handled hands consistently. None could maintain identity across a series — generating a character in one composition and the same character in another remained an open problem. None could produce print-resolution output by default. None handled CMYK color. None had a brand-consistency story beyond what the user could prompt their way to.
Where they differed was in everything around the model: the working interface, the deployment story, the asset-management story, the rights story, the production-pipeline integration. Brand teams that succeeded with AI imagery in 2023 were the teams that picked the tool whose surrounding properties matched their actual workflow, not the tool that produced the most attractive individual outputs in a marketing demonstration.
Where this left the market
By the end of 2023, the brand creative tooling market had begun to bifurcate. The teams running serious production volume had built workflows around Stable Diffusion’s deployment flexibility and control depth, even at the cost of operational complexity. The teams running atmospheric or editorial work had built workflows around Midjourney’s output quality, even at the cost of pipeline friction. The teams running stakeholder-alignment and concept work had standardized on DALL·E for its accessibility.
What was largely missing from the market was a tool that combined production-grade output with production-grade workflow integration, trained on a corpus appropriate to brand creative work specifically. The next year would be defined by attempts — from existing players and from new entrants — to close that gap.
For the broader argument about training corpora that drove that next phase, see AI Creative vs. AI Slop.