Skip to content

Firefly Graph Arrives a Year Late, and the Capsule Is the Part Worth Stealing

Adobe opened Firefly Graph to individual accounts on 10 November as a public beta: 250-plus nodes, partner models and Photoshop operations on one canvas. The canvas is the familiar half. The capsule, which packages a graph into a tool a colleague can run without reading it, is the part that changes how a team hands work over.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
Five small brass and glass instruments arranged left to right on dark felt, linked by short curved rods, with a three-dial hand controller resting in front of them
Editorial illustration about chained steps and exposed controls; not a product screenshot or a model output.

On 10 November, in the middle of Adobe MAX, Firefly Graph stopped being something you read about and started being something you could open. The Japanese studio CreativeEdge noticed the same day and wrote it up that evening: the tool shown as Project Graph at last year's MAX, shipped to enterprise accounts around June as Adobe Graph, had finally been handed to individual users. Their verdict on the interface is the useful one. If you have used Figma Weave, you will pick it up quickly. If you have used ComfyUI, it will feel like home.

That ordering tells you most of what you need to know. Node canvases are not new. What is new is where this one lives, and what it lets you hand to someone else.

I have not run the beta. Everything below comes from Adobe's release, its help pages, and one thorough Japanese first-day report, so treat the capability claims as Adobe's own and the judgment as mine.

What shipped, and what only looks shipped

Adobe's announcement mixes four availability tiers, which is worth untangling before anyone plans a week around it:

  • Generally available now: Edit Text, Retouch, Relight, and the Quick Actions gallery.
  • Public beta: Firefly Graph and Firefly Design.
  • Private beta by waitlist: Firefly Builder, which turns a prompt into an interactive prototype.
  • Beta labelling without a separate gate: the Firefly AI Assistant side of the expanded Creative Agent.

The release talks about all of it in one breath, and the breath is persuasive. The tiers are not. A workflow you sell to a client on a public beta is a workflow you may be rebuilding in February, and Firefly Builder is not something you can plan for at all until an invite arrives.

CreativeEdge adds two details the marketing does not. The interface redesign that turned Firefly into an "AI-first" product has been rolling out for a month, and Firefly Graph's UI is still English-only even though Adobe's tutorials for it are localised to Japanese. A tool that is documented in your language and operated in another one has been pushed out fast. That is not a complaint, just a marker of maturity.

The canvas is the least interesting part

Adobe's help pages describe a library of more than 250 nodes: text-to-image generation on Adobe's own models plus partner models (Flux, Gemini, GPT Image, Imagen, Ideogram, Runway, with Kling and Seedance on the video side), Creative Cloud editing operations such as masking, hue shift, gaussian blur, colour grading and compositing, and utility nodes for resizing, layer blending and picking an item out of a list. Shader nodes are in there too, and fuller video and 3D support is listed as planned.

If you have built anything in ComfyUI, that list is familiar to the point of being unremarkable. The difference is not the shape of the canvas, it is what the nodes are made of. ComfyUI nodes wrap models and image operations that the community wrote. Firefly Graph nodes wrap the same operations a designer already performs in Photoshop, Illustrator and Premiere, and Adobe is explicit that the product exists for "upstream creative work" like designing, authoring and templating workflows.

That matters more than model access, which anyone can resell. The expensive part of a real production pipeline is not generating the image, it is the twelve steps around it: the crop, the grade that matches the previous campaign, the layer stack that the printer accepts. A tool that can call those steps as nodes is doing something a model marketplace cannot.

The capsule is the idea to take

Here is the feature I would steal if I were building a workflow tool tomorrow. You select a group of nodes and collapse them into a subgraph, which exposes only the inputs and outputs you designate and hides everything inside. Package that subgraph with its exposed controls and you get what Adobe calls a capsule: a self-contained tool that a teammate can run without understanding how it was built. One correction before anyone plans a week around it: the capsule workflow is documented under Firefly Creative Production for Enterprise and Firefly Graph for Enterprise, which Adobe sells with Creative Cloud for Enterprise Edition 5. Collapsing nodes in the individual beta gives you a subgraph, not a distributable capsule.

The mechanics underneath are ordinary and good. Input controls are standalone nodes that feed sliders, angle dials and colour pickers into hard-coded values, so the person running the graph turns three dials instead of editing fifteen settings. Lists pass along double-line connections when a node returns several results, and a "get item from list" node chooses which one continues downstream. Sharing is view-only or edit access.

None of that is new in the abstract. It is how Houdini digital assets and Nuke gizmos have worked for years, and ComfyUI users approximate it by handing over a JSON file with a paragraph of instructions attached. What is new is the audience. A capsule can be handed to a designer who has no interest in node graphs, and the handoff does not require them to install anything, learn a package manager, or trust a workflow file from a stranger on the internet. That audience is currently defined by an enterprise contract rather than by an Adobe account, which narrows where the idea applies today without narrowing the idea.

That is a real answer to the problem that has kept node tools inside a small priesthood. The author and the operator can be different people with different skills, which is how every other production department already works.

The assistant builds the graph. You still press Run.

Adobe's guide for the in-graph AI Assistant is unusually blunt about what it will not do, and the list is the most honest thing in the release:

  • It can create and modify workflows, but cannot run them. You start execution yourself.
  • It cannot add input assets or interpret their contents.
  • You cannot attach files to its chat.
  • Previous conversations are not accessible inside the graph.

Read that as the boundary between a saved process and a delegated job. The assistant will assemble the machine and explain which node blends two images, and then it stops and waits for a human with the source files. Whether that is a deliberate design choice or a beta limitation, the effect on your week is identical: this is an authoring aid, not an operator.

That distinction is the reason I would not file Firefly Graph under "agents". It is closer to the opposite. We argued elsewhere on this blog that agent-driven pipelines are the more practical interface for repeatable creative work, because a saved graph still needs someone who knows what to press. Firefly Graph is the strongest counter-case I have seen, and it makes the counter-case by admitting the limit out loud instead of hiding it behind an autopilot button.

Credits are a property of the graph now

Adobe's product FAQ is careful here: not every node costs credits, nodes that use generative models do, the number varies by model and node, and the cost is displayed on the node before you generate. The assistant documentation repeats the warning in its own words, telling you to review AI-built workflows before running them "especially when actions consume generative credits", and points at a credits icon that reports what an action actually used.

For anyone who has planned a shoot around a credit budget, the interesting shift is where the uncertainty moves. A single generation has a price you can look up. A graph that branches, or iterates over a list of twelve product variants, or chains an LLM rewrite into three regeneration attempts, has a price that depends on how many times it runs. Budgeting moves from the render to the run. If the graph is going to be used by people who did not build it, the number they need is not the cost of one image, it is the cost of one complete pass with their inputs, measured rather than estimated. Treat that measurement as a sample rather than a cap when the workflow is marked "Credits vary", since a node that returns several results or loops over a list can cost more or less on the next run with different inputs. Our credit-first planning for a Firefly anime short makes the same argument from the other end: decide what the expensive step may cost before you get attached to the output.

Where the local stack still wins

Three things keep ComfyUI and its neighbours relevant after this launch.

Ecosystem breadth. Adobe's 250-plus nodes are curated. ComfyUI's node repositories are not curated at all, which is why they cover things Adobe will never ship, and why half of them break after an update. Curated wins on reliability, open wins on coverage.

Ownership. Weights you downloaded do not change pricing, do not require a plan tier, and keep working offline. They come with licence problems of their own, and I have written about a model whose licence excludes four major markets precisely because "free download" is not the same as "free to ship". But the trade is legible, and the ceiling is a graphics card rather than a monthly allocation.

Scale-out. Adobe scopes Graph to workflow authoring and puts high-volume execution of published workflows in a separate enterprise product, Firefly Creative Production. So the graph is not the factory. If your job is 4,000 assets a week, this is the tool that designs the process, not the one that runs it.

What I would build first

Pick one repeating task with a handoff attached, meaning someone other than you has to produce it. A campaign set is the obvious candidate: forty product shots that need the same backdrop, the same crop, the same three export sizes.

Then build the graph and immediately collapse it. Three controls is the ceiling, not the target. If the operator needs to choose between four type treatments and two lighting directions, that is eight branches to test and eight ways to ship the wrong one.

Measure a full pass before you promise a cost. One run, real inputs, credits icon checked at each generative node. Then write the number down where the person running it can see it.

Version by duplicating and dating. With public beta status and view or edit sharing, the graph that worked last Monday can be different on Thursday. A copy named for the campaign is cheaper than an argument about who changed what.

And keep the exit. The output spec, meaning sizes, crops and naming, should live somewhere that is not the graph. That way the day Adobe moves a node, or the day the beta closes, you can rebuild the pipeline in something else without renegotiating what it was supposed to produce.

Firefly Graph is a year late and arrives in beta, and neither of those is the story. The story is that a subscription design suite has finally shipped the part of node tooling that was missing: a way to give a colleague a tool instead of a diagram. Beta or not, that idea is going to be copied.

Sources

Share this guide
X LinkedIn