Anatomy of a ComfyUI workflow, and the four nodes that matter

Most local generation starts by downloading someone else's graph. It opens as a sprawl of boxes and wires, most of which are plumbing, and the temptation is either to run it blindly or to give up.

The useful realisation is that almost every image workflow has the same skeleton underneath the sprawl. Once you can see the skeleton, the forty nodes resolve into five or six that matter and a great deal of wiring between them.

Learning to read the skeleton takes half an hour and makes every workflow you ever download adjustable.

1. Load the model. One node bringing in the weights.
2. Encode the prompts. Text into a form the sampler can use — usually two nodes, positive and negative.
3. Create the latent. An empty canvas at your chosen dimensions.
4. Sample. The step that does the actual generating, taking model, prompts and latent.
5. Decode. Latent back into an image.
6. Save. Write it out.

Every extra node in a complex graph attaches to one of these six. Find them first and the rest becomes readable as decoration around a known structure.

Prompt — the text nodes. Two of them; make sure you are editing the positive one, which is the one wired to the sampler's positive input.

Dimensions — on the empty latent node. Changing this changes memory use sharply, not linearly.

Seed — on the sampler. Fix it while testing other variables; randomise it when you want variation.

Steps — also on the sampler, and the one most often wrong for the model in use. Accelerated models want far fewer steps than the graph's default usually specifies.

Everything else can be left alone until you have a specific reason.

Graphs published online were generally built on discrete-card machines and specify a loader for a format that suits them. On Apple Silicon that is usually the wrong choice and the fix is to swap the loader node for one appropriate to the format you are actually running.

Second thing to check: the sampler's step count and scheduler against what your model expects. A graph built for a standard model applied to an accelerated one, at the original step count, produces slow and worse results.

Third: any conditional or switch nodes. These sometimes fail validation on newer versions, and removing them and wiring the branch you want directly is quicker than debugging why the graph refuses to run.

When output is wrong, attach a preview to the decode stage and confirm the basic path works before investigating anything elaborate.

If the simple path is correct, re-enable the additional stages one at a time. The stage where the result goes wrong is the one to look at, and this narrows a forty-node graph to one region in a few runs.

Bypass rather than delete while testing. Most nodes can be disabled without unwiring them, which keeps the graph intact and makes the comparison honest.

Once a downloaded graph works, save it under your own name with a note in the graph itself: what it is for, what you changed, what to adjust.

Many workflows carry their author's notes inside note nodes — trigger words, model links, warnings. Read those before assuming a setting is arbitrary; they frequently explain the one non-obvious choice.

Over time you accumulate a small set of graphs you understand completely, which is worth more than a large collection of downloads you run without knowing why they work.