← All classes
CLASS 8June 17 · Node-Based Workflows in Krea

"Krea is Microsoft Paint.Comfy UI is free Photoshop."

Seb Saga
Taught bySeb SagaSebastian Sagastume · Founder, Seb Saga LLC
Pratt After Effects & video editing instructor since 2022

The class in brief

Guest Seb Saga (video production and animation, clients including Disney and the NY Islanders) taught node-based AI workflows in Krea: chaining single-purpose blocks into a reusable pipeline instead of re-typing one prompt at a time. Six archetypes, a live build from blank canvas to finished video, and a blunt account of where the tool breaks. After this page you can name the right archetype for a job, let an agent build a workflow's bones from one sentence, and know when to walk away from Krea for ComfyUI instead.

The night at a glance

This class moves. Watch the four recovered demo clips with audio →

Why this matters · 6:04 PM

A node does one job. Chain enough of them and the workflow runs itself.

Seb's core thesis, opening the night : nodes are "smart LEGO blocks," each one generating, editing, upscaling, or animating, and chained together they become a reusable pipeline. "It's basically programming, but with visual blocks." What the room would build that night fell into three shapes: scalable product placement, a video chain, and multi-layer compositing .

Before reaching for a node canvas at all, Seb gave six questions to ask first : does it need to scale, can it be reused or handed off, are you comparing models, is it repetitive, do you need several models orchestrated together, or do you need absolute control (the asterisk on that last one: then the real answer is ComfyUI) .

6
questions to ask before you build a node workflow at all. Answer no to all six, and one prompt still does the job.
3
things the room would build that night: product placement, a video chain, and multi-layer compositing.

The framework · 6:17 PM

Six ways to build a workflow

01

Lab benchcompare models or settings

Same prompt, several models, side by side. Seb's test: a skateboard trick, where one model nailed it and others missed entirely. Also test aspect ratios; models perform differently across them, a training-data artifact.

02

Factoryone input, same process out, at scale

Multi-angle character generation, turnaround sheets, a product photo becoming a model wearing it plus a 360 view. The same output shape, run over and over.

03

Multiplierone asset, many scenarios

A single product shot placed into several environments at once: the product-placement pattern the night's live build used.

04

Relayoutput feeds the next model

Image into edit into video, one model's output becoming the next model's input, down the line.

05

Compositorassemble the pieces, then unify

Generate scene layers separately, remove backgrounds, assemble, then run the flat composite through an editing model to match shadows and light.

06

Control rigControlNets, structural prompting

Pose skeletons and depth maps as a third type of prompt, alongside text and images. This is ComfyUI territory, not Krea's.

The Factory, product photo to turnaround set
The Factory, product photo to turnaround set · 6:21 PM

The exercise · 6:52 PM

Building the first chain live

The build

Blank canvas to a finished squirrel-in-a-tuxedo video.

Seb built the first chain from nothing: a Z-Image node with a plain sunset prompt , then a Text node feeding an LLM Call node feeding the image node, three blocks long . From there: a Qwen edit put the squirrel in a tuxedo and top hat, and an LTX-2 node turned the result into a short video.

"The Krea agent builds the bones. You still write the prompts that matter."

The preset that saves the step

Seb's LLM-node system prompt, saved and reused across every workflow he builds: "Expand on the user's image prompt. Ensure to describe settings, tone, feel, angle, color palette, and any other descriptions when relevant. Respond with only the detailed prompt." One block, written once.

The agent builds the bones

Asked in plain language to place a sweater into four environments, the Krea agent built the prompt-writer, the splitter, and four generation nodes on its own . Scaled up to a cardigan and eight environments, the pattern held .

Cheap first, then swap

The Photoshoot App demo ran a skateboard product photo through cheap models first, four style variations for the cost of testing, before anyone touched a frontier model .

The craft · 7:45 PM

Assemble the pieces separately, then let one model unify the light.

Post-break, the Compositor archetype got its full demo: a fox, a stone bowl, and a zen-garden background generated as three separate layers, backgrounds removed, then handed to a single editing model with one instruction, integrate the layers with realistic shadows and light .

The video chain closed the demo block: a subway phone ad built from a green-screened phone, three Seedance clips stitched end to end, and a 2x upscale on the way out .

Subway ad video chain
The subway ad, stitched from three clips · 8:05 PM

Methods and prompts

Five methods to take with you

METHOD 01 · THE SIX QUESTIONS part of: keep the judgment human

Decide if you need nodes at all

Before building anything, answer the six diagnostic questions yourself. A single prompt still does the job for most tasks; nodes earn their complexity only when the answer is genuinely yes.

Where it came fromTaught by guest Sebastian Sagastume, known as Seb Saga, at the start of class 8, as six diagnostic questions to ask before building any node workflow in Krea.Use it whenUse this before opening a node canvas, when you are not sure whether a task actually needs a multi-step workflow or a single prompt would do.
Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Here's the task I'm about to automate: [describe it]. I'll answer these six questions myself first, then you check me: does it need to scale past a handful of outputs, could it be reused or handed off, am I comparing models or settings, is it repetitive, does it need several models orchestrated together, do I need pixel-level control? My answers: [yours]. Tell me honestly whether I need a node workflow, or a single prompt still does the job.

You will know it worked whenit tells you plainly whether a single prompt still covers the task, based on your own six answers, not a default push toward nodes.

METHOD 02 · THE AGENT part of: build tools not chats

Let the agent build the bones

Describe the outcome in plain language and let the agent lay out the nodes; you review and fix the prompts it writes, rather than wiring every connection by hand.

Where it came fromDemonstrated by Seb Saga in class 8's live build, where he described an outcome in plain language and let Krea's built-in agent lay out the node structure before reviewing the prompts it wrote.Use it whenUse this when starting a new node workflow and you would rather describe the goal in words than wire every node connection by hand.
Same idea elsewhereStand up a reusable Project or Gem (class 4) · Weight your references on purpose (class 5) · Let one model write the prompt for another (class 6)

Working prompt

I want a node workflow that [describe the outcome, e.g. "places this product photo into 4 lifestyle environments"]. Build it: the input node, a prompt-writer step, and one generation node per output. Use the cheapest capable model for a first pass. Show me the prompts you wrote before I run it.

You will know it worked whenit shows you the prompts it wrote for each node before running anything, so you can catch a bad one before it generates.

METHOD 03 · THE PRESET part of: context beats prompts

Save a reusable prompt-enhancer preset

Seb's own system-prompt block, saved once and applied to any image node. One saved instruction, reused across every workflow instead of rewritten each time.

Where it came fromSeb Saga's own saved system prompt, shown in class 8 as a reusable LLM node instruction he applies to any image-generating node instead of rewriting it each time.Use it whenSave this once you find an instruction that reliably improves your image prompts, so every future node workflow can reuse it without retyping.
Same idea elsewhereControl the input and build gradually (class 2) · Show it your voice and not just the output (class 3) · Evolve the reference instead of restarting (class 5)

Working prompt

Expand on the user's image prompt. Ensure to describe settings, tone, feel, angle, color palette, and any other descriptions when relevant. Respond with only the detailed prompt.

You will know it worked whenevery image node using this preset expands your short prompt into settings, tone, angle, and color palette without you typing them each time.

METHOD 04 · COST DISCIPLINE part of: reject the first draft

Build cheap and then swap the model up

Verify the whole chain works on the cheapest model that can do the job, then upgrade only the one node that needs it. A two-click dropdown change, not a full rebuild.

Where it came fromSeb Saga's cost-discipline habit, taught in class 8: build and verify a full node chain on the cheapest capable model, then upgrade only the one node that actually needs a stronger model.Use it whenUse this when setting up a new workflow and you want to control cost, testing the whole chain cheaply before spending on a frontier model.
Same idea elsewhereSet up a debate partner (class 4) · Critique mode (class 10) · Find where the workflow breaks (class 11)

Working prompt

Build this workflow with [name the cheapest capable model, e.g. Z-Image]. Once every node runs clean end to end, tell me which single node would benefit most from a frontier model, and swap only that one.

You will know it worked whenit names exactly one node worth upgrading to a stronger model, not a blanket recommendation to swap the whole workflow.

METHOD 05 · THE 2X2 TRICK part of: generate multiples

The 2x2 grid trick

Ask for four variations in one grid instead of four separate generations. Better scene consistency, and four options for one generation's cost.

Where it came fromShared by Seb Saga in class 8 as a way to get four scene variations from a single generation instead of four separate ones, with better consistency between the results.Use it whenUse this when you want several options from one image generation instead of paying for and comparing four separate runs.
Same idea elsewhereRefine with constraints (class 9)

Working prompt

Generate a 2x2 grid showing four variations of this scene: [describe it]. Keep the composition and framing consistent across all four so I can crop out whichever one wins.

You will know it worked whenthe four variations share the same composition and framing, so you can crop any one out without the scene shifting underneath it.

Where it broke · all night

The line splitter and the honest downsides slide

Seb put Krea's weaknesses on screen before anyone hit a bug: buggy, less scalable, credit-hungry, less control than the alternatives . The line splitter, a utility that fans one multi-line output into several prompts, was the night's biggest live failure: two workflows silently produced nothing, and the fix was asking the agent to abandon the splitter and write a prompt into each node by hand .

Related classes

Where this class connects

The shelf

Tools and references

Tools that night

  • krea.ai/nodes, plus the built-in Krea agent (the "K" button)
  • ComfyUI (comfy.org), local and free; Comfy Cloud, ~$20/month hosted
  • Named as peer node platforms: Runway nodes, Weavy, Flora
  • Cheap tier: Z-Image, LTX2, Qwen Image Edit, Gemini 2.5 Flash. Frontier tier: Nano Banana Pro, Seedance, Claude Sonnet/Opus
  • Topaz, Magnific (upscalers); Photopea (free Photoshop-alike, compositor round-trips)

Named in the room

  • Seb Saga: sebsaga.com · hi@sebsaga.com, open to questions and connections
  • The skateboard "Benihana" trick, the Lab Bench's test case across models
  • This page carries the frames that teach the lesson; the full run of stills from the recording is not included.
← Class 7 · AI-Enhanced Workflows CLASS 8 OF 24 Class 9 → AI Remix Jam