ComfyUI vs Self-Hosted Stable Diffusion and Flux vs Adobe Firefly vs Midjourney: What You're Committing to Own

Generative imagery usually enters a company through one person. A designer starts using Midjourney, a marketer tries Firefly, an engineer builds a ComfyUI workflow on a spare GPU, and within a year product photos, catalog variations, and campaign assets depend on tools nobody approved, licensed, or documented.

Choosing a tool takes a week. Owning a production pipeline takes years. This piece compares four common routes (ComfyUI, self-hosted Stable Diffusion and Flux models, Adobe Firefly, and Midjourney) on the questions that matter once the images are part of how the business operates: what you are licensed to do, who runs the pipeline, what it costs to run, and what it would take to leave.

Four different kinds of thing

As with most technology comparisons, these are not four versions of the same product.

ComfyUI is an open-source, node-based interface and execution engine for diffusion models, released under the GPL-3.0 license. You build a workflow as a graph (load a model, apply a prompt, condition on an image, upscale, save) and save it as a file that can be versioned and rerun. It does not include a model. It runs the models you give it, on hardware you provide, and it is extended through a large ecosystem of community custom nodes.

Self-hosted Stable Diffusion and Flux are the model weights you run inside ComfyUI or another runtime. Their licenses differ by model and matter commercially. Stability AI’s Community License allows free use, including commercial use, for organizations with less than $1 million in annual revenue. Above that, Stability requires an Enterprise License. Black Forest Labs released FLUX.1 [schnell] under the Apache 2.0 license, while FLUX.1 [dev] carries a non-commercial license, with commercial use licensed separately. Check the license for the exact model you run, not the model family.

Adobe Firefly is Adobe’s family of generative models, available in Adobe’s applications and through APIs. Adobe says its models are trained on licensed content, including Adobe Stock, and public domain content. It offers IP indemnification for generated content on qualifying enterprise plans, with terms, and that coverage does not extend to non-Adobe partner models offered inside Adobe’s tools. Firefly outputs carry Content Credentials, the C2PA provenance metadata.

Midjourney is a hosted image generation service used through its website and Discord. At the time of writing it does not offer an official public API, and its terms prohibit automated access, so the third-party “Midjourney APIs” on the market are unofficial wrappers that put the account at risk.

Start with the capability, not the tool

The architecture question is which business capability the imagery supports, and how much the business depends on it. A team generating mood boards for internal discussion has a very different requirement from a retailer generating thousands of catalog variations a month, or a manufacturer rendering product configurations for a quoting tool.

Map it the way you would any other capability: name it (Product Content Production, Campaign Asset Creation, Configurator Rendering), give it an owner, and define what good looks like. Then list what you will own regardless of the tool:

Question 1: Who runs it after launch?

Midjourney and Firefly are run by their vendors. Your operational work is accounts, seats, governance, and review, plus integration if you use Firefly’s APIs. Midjourney’s lack of an official API keeps it a tool for people, not a component in an automated pipeline.

ComfyUI with self-hosted models is the opposite. You run all of it: the GPU hosting, the runtime, the models, and the workflows. That gives you control over consistency, cost per image at volume, and data that never leaves your environment. It also gives you a real application to own:

Question 2: What does it cost to run?

Hosted tools charge per seat, per plan, or per generation. Costs are easy to see and grow with headcount and usage. The hidden cost is uncontrolled use: assets created under personal accounts, on personal plans, with terms nobody reviewed.

Self-hosting moves the cost to infrastructure and people. At low volume it is almost always more expensive than a hosted plan, once you count the engineering time. At high, steady volume, with a fixed workflow, the cost per image can fall well below hosted pricing. The deciding factor is usually not the GPU bill but whether you have someone to own the pipeline.

Licensing is the cost that surprises people. A company that grows past Stability AI’s revenue threshold, or that moved a FLUX.1 [dev] workflow from a test into production, may already be outside its license.

Question 3: What would it take to leave?

Workflows in ComfyUI are portable in format (they are files), but they are bound to specific models and nodes. Moving to a different model usually means re-tuning the workflow and re-approving the output style.

Hosted tools hold your history. Prompts, generations, and settings may live only in the vendor’s interface. Export what you need to keep, and keep the approved assets and their provenance in your own digital asset management system.

Brand consistency is the real lock-in. If your catalog has a look that depends on one model’s behavior, switching models changes the look. That is a business decision, not a technical one, and it deserves to be made deliberately.

A practical way to decide

  1. Inventory what is in use today, including personal accounts, and which published assets came from which tool.
  2. Name the capability and its owner, and decide how dependent the business is on it.
  3. Match the tool to the requirement. Indemnified, vendor-run generation (Firefly) suits marketing teams that need rights protection. A hosted creative tool (Midjourney) suits ideation by people. A self-hosted pipeline (ComfyUI with properly licensed models) suits high-volume, consistent, automated production where you can staff the ownership.
  4. Check every model license against your revenue and use, and record the result.
  5. Govern the output: review, provenance, and storage of what you publish.

For the governance policy itself, the NIST AI Risk Management Framework and ISO/IEC 42001 give you a structure. Neither replaces knowing which model made which image under which license.

Where we come in

Generative imagery becomes a business system the moment a product page, catalog, or quoting tool depends on it. That is usually also the moment the person who built the workflow moves on.

We take on that ownership: after-build ownership of self-hosted pipelines and the applications that call them, with pinned versions, reviewed custom nodes, monitoring, licensing records, and documentation. We run change management so model and workflow updates are tested against approved outputs before they reach production. And when a pipeline the business depends on stops working, or a license question surfaces, we step in urgently and leave it supportable.

If your team is already producing imagery this way, tell us about it.


Tell us what you're running.

Two quick steps. I'll read it personally and reply with an honest read on your situation and whether our team can help, usually within one business day.

  • I lead every engagement personally
  • No sales sequences, and your details are only used to reply

Prefer email? [email protected]

Step 1 of 2

Which sounds most like you?
What's prompting this? Pick any that apply
How many business-critical custom systems?
Timeline
Where should I reply?