ALL NODES LEAD HERE

From node-novice to node-native: a working guide to AI workflow systems for creatives

Table of Contents

1. The Grammar Nobody Explains

2. The Three Node Types That Actually Matter

3. The Capability Gap Nobody Warns You About

4. Reference Stacking and Why You're Leaving Power on the Table

5. Where to Start (and Where Not To)

6. Notable Nodes Worth Knowing About

7. One More Thing: Adjacent Tools Worth Watching

1. The Grammar Nobody Explains

If you've spent any time in Houdini, Cinema 4D's node material editor, Nuke, or Blender's shader tree, your first look at an AI node board will feel basic. Shallow branching, limited data types, none of the technical depth you're used to. The instinct is dismissal.

That's the wrong read.

AI node systems aren't trying to be compositing pipelines. They're trying to give creative people repeatable, controllable workflows without requiring them to write code. Judged on that basis, the better platforms are doing something genuinely useful.

For everyone who's never touched a node graph before: the thing that looks intimidating isn't. Every node system, everywhere, runs on the same grammar. A node is a box. It takes something in on the left. It does something to it. It passes something changed out on the right. That's it. Chain enough boxes together and you have a workflow.

If flowcharts make sense to you, you're already most of the way there. A node board is essentially a flowchart—but without the yes/no/maybe branching. There are no decision diamonds, no conditional forks. Data flows in one direction, through a chain of operations, toward an output. That's it. If anything, it's a simpler mechanism than a flowchart. The complexity you see on an advanced node board isn't structural—it's just more steps in the chain.

One rule before anything else: connections only work between matching types. Image wires connect to image inputs. Text connects to text. Video to video. The system won't let you connect incompatible types. Follow the wires and you won't get lost.

2. The Three Node Types That Actually Matter

Not all nodes are equal, and the platforms that give you genuine creative control are the ones that offer all three of the following types. Most confusion in node systems comes from not knowing which type of node you're looking at, or why it's there.

Input nodes

Input nodes are where everything begins. They're the sources—the raw material the rest of the workflow operates on. On most platforms you'll find these colour-coded to help you track what type of data is flowing where.

Common input node types across platforms:

Image nodes — a single static image fed into the workflow. The most basic reference type and the most universally supported.

Video nodes — a video clip as input. Enables frame extraction, motion reference, start/end frame control, and video-to-video workflows.

Text and prompt nodes — the written instruction passed to a generative model. Can be a simple prompt or a structured input built by an LLM node earlier in the chain.

Mask nodes — defines a region of an image for targeted operations. Inpainting, selective editing, and compositing all depend on a clean mask input.

LoRA nodes — a trained style or subject model loaded as a reference input. Connects to a generative node to influence output character, not just through prompting but through trained weighting.

Style reference nodes — visual inputs that guide aesthetic direction without full LoRA training. Connects an image to a generative model as a style anchor.

Audio nodes — audio file input for lipsync, sound-driven generation, or audio-to-video workflows.

Array and list nodes — found in Flora and Krea. Enable multiple inputs of the same type to be grouped and passed together, which is the foundation for batch processing.

It's also worth understanding that input nodes aren't used in isolation. A single generative node will often accept multiple input types simultaneously—and combining them is where the real control begins.

The most familiar version of this is positive and negative text prompts wired into the same generation node: tell the model what you want and what you don't want in the same pass. Start frame and end frame nodes fed into a video model are the same principle applied to motion—you're defining where the sequence begins and where it lands, with the model responsible for what happens between.

But it goes further than that. On the more capable platforms you can wire image references, a LoRA, a video, a 3D asset, and a text prompt into a single generative node at the same time. Each input is contributing something different—subject, style, motion, geometry, instruction—and the model is reconciling all of it simultaneously. This is the difference between prompting and directing. You're not describing what you want in words and hoping the model interprets correctly. You're providing evidence from multiple directions and letting the model find the intersection.

The key thing to understand about input nodes is that they set the ceiling. A generative model can only work with what you give it. More input types, more connections, more reference fidelity—the output has more to anchor to.

Output and generative nodes

These are the nodes that cost credits. They're the ones with a Run button. They take everything flowing in from the left, process it through a model, and produce something new on the right.

Common generative output node types across platforms:

Generate image nodes — text-to-image and image-to-image generation. The most common node type on any platform. Each model available on the platform typically has its own generate image node, with different input limits, reference image caps, and style capabilities.

Edit image nodes — inpainting, outpainting, prompt-based editing of an existing image. Works on what's already there rather than generating from scratch.

Enhance and upscale nodes — post-generative quality improvement. Increase resolution, add detail, fix degraded output. Krea offers options including Topaz up to 22K resolution. These nodes sit after generation in the workflow, not before.

Generate video nodes — image-to-video, text-to-video, or reference-to-video. Start frame, end frame, camera control, and duration are typical parameters. Model choice here matters significantly—different video models have very different motion characteristics.

Generate 3D nodes — image or text to 3D asset. Available on Krea via TRELLIS and Tripo. Not yet common across all platforms but growing.

Audio generation nodes — text-to-speech, sound design, and synchronised audio for video. ElevenLabs, ThinkSound, and MMAudio are available as distinct nodes within Krea's system rather than as separate tools you'd switch to.

Lipsync nodes — drive lip movement on a face using an audio input. Fabric and Hedra on Krea. Increasingly relevant for video production workflows that need to stay inside a single pipeline.

One important note on generative nodes: the model inside the node and the parameters the platform exposes are two different things. Always check what inputs the platform actually surfaces versus what the underlying model supports. The capability gap covered in section 3 lives here.

Tool and edit nodes — the layer that matters most

This is the middle. The layer between what you put in and what comes out. And it's the most important layer for actual creative work, because it's where you exercise control rather than hope.

If a platform's node system only gives you input nodes and generative nodes, you have a workflow. If it gives you a rich middle layer, you have a creative pipeline. The distinction is significant.

These nodes are typically non-generative—they don't cost credits, they run instantly, and they operate on data rather than generating new content. That makes them fast to iterate and free to experiment with.

Common tool and edit node types across platforms, and what they're actually for:

Compositor node (Krea, Weavy) — layers multiple images or videos together with blend modes, opacity control, position, scale, and rotation. Up to 8 layers in Krea. This is your in-workflow compositing stage. Use it to combine generated elements before they reach the next generative node, or to merge outputs at the end of the pipeline.

Painter and mask editor nodes (Krea, Weavy) — draw directly onto a canvas to create hand-painted masks or rough sketches that feed into sketch-to-image models. For anyone coming from a drawing background this is the most direct creative input available in a node system. Your marks become the reference.

Background removal node (Krea) — AI-powered subject isolation. Removes the background from an image output before it continues through the workflow. Particularly useful in product photography pipelines where the subject needs to be composited into a new environment downstream.

Levels node (Weavy) — adjusts brightness, contrast, and tonal range using shadow, midtone, and highlight controls. Use it to correct exposure on a reference image before passing it to a generative model, or to grade an output before export.

Blur node (Krea, Weavy) — Gaussian or box blur with adjustable intensity. Useful for softening masks, creating depth-of-field references, or deliberately reducing detail in a reference to give a generative model more interpretive freedom.

Crop and resize nodes (Krea, Weavy) — prepare inputs to match model requirements, reframe outputs, or adapt aspect ratio for different formats. Simple but essential for keeping a workflow clean rather than fixing mismatched dimensions manually at every step.

Invert node (Krea, Weavy) — flips pixel values. Primarily used for mask inversion, where the area you want to affect and the area you want to protect need to swap. Also useful for creative colour effects.

Channels node (Weavy) — accesses individual RGB and Alpha channels from an image or video. Advanced compositing use. If you're coming from a VFX background you'll know exactly what this is for; if you're not, it's worth knowing it exists for the day you need it.

Extract video frame / Get video frame (Krea, Weavy) — pulls a specific frame from a video as a static image. Use it to grab a start or end frame to feed into an image-to-video model, or to extract a reference frame from footage you want to match.

LLM Call node (Krea, Weavy, Flora) — places a language model mid-workflow. Analyse an image, generate a structured prompt from a brief, build a shot list, rewrite a prompt to better match a model's preferred syntax. The LLM isn't just at the front of the pipeline—it can sit anywhere in it, processing text or analysing visual inputs before passing results downstream.

Text utility nodes — concatenation, line splitting, text overlay. Build prompt variants programmatically, combine multiple text inputs into a single prompt, or add watermarks and captions to output images without leaving the workflow.

Motion transfer nodes (Krea) — capture motion from a source video and apply it to a different subject. The movement pattern becomes an input, independent of the character performing it.

Video Time Ramp node (Krea) — remap video timing using curve types including bezier, ease-in, and ease-out. Pacing as a pipeline value rather than a post-production decision.

Video utility nodes (Krea) — trim, crop, adjust speed, stitch multiple clips, combine video with audio, adjust hue and saturation. A working edit suite inside the node board.

Router node (Flora) — groups multiple assets into a single node to keep complex boards legible. Workflow hygiene as a named node rather than just a canvas organisation feature.

The consistent principle across all of these: they let you shape data before it reaches a generative model, and shape output before it leaves the pipeline. Every one of them reduces the gap between what you intended and what gets produced.

3. The Capability Gap Nobody Warns You About

This is the thing that catches people out.

The model's ceiling and the platform's ceiling are not the same number.

A model might natively support up to 15 reference images. The node board you're using might only expose 3. You're not using the model's full capability—you're using whatever the platform chose to implement. And the gap is often significant.

Krea's own documentation illustrates this clearly. Nano Banana supports 3 image references as a node input. Nano Banana Pro supports 15. Flux sits at 2. ChatGPT Image 1.5 supports 15. Same platform, same board, wildly different input ceilings depending on which model you're wired to.

Before you conclude that a workflow isn't working, check what you're actually passing in. The problem is often not the model.

4. Reference Stacking and Why You're Leaving Power on the Table

Related to the above: you don't need a full training dataset to get meaningful style or subject consistency.

The common assumption is that proper LoRA training requires substantial data volume, and without it you're not going to get reliable results. That assumption leads people to either over-invest in training or give up on reference-based control entirely.

Neither is necessary. Six reference images, or a short video paired with a defined start and end frame, gives a model significantly more to anchor to than a single reference. It won't replace a properly trained LoRA for production work, but for exploration, rapid iteration, or situations where a full training run isn't justified, it's a legitimate creative input—not a consolation prize.

Max out what your platform exposes. Then check whether the model supports more than the platform is showing you.

5. Where to Start (and Where Not To)

Three platforms that are genuinely good entry points, each with a different emphasis:

Krea is the most full-featured of the three. Rich node taxonomy covering generation, editing, enhancement, motion transfer, lipsync, audio, and 3D. The capability gap evidence earlier in this article came from their own documentation. If you want breadth, start here.

Weavy (Figma Weave) has the clearest structural logic for newcomers. It splits nodes explicitly into two categories: generative nodes, which cost credits and have a Run button, and non-generative nodes, which are free and instant. That single distinction removes a lot of confusion early on. The Painter node—hand-drawn masks feeding into sketch-to-image models—is worth exploring as a direct creative input method. If Figma is already your primary creative tool, there's an additional reason to get familiar with Weavy specifically: Figma acquired it, and it will likely begin appearing more deeply inside Figma over time. Getting comfortable with it now isn't just useful—it's probably inevitable.

Flora is where the category gets genuinely interesting for production work. The Batch Node changes the equation entirely. Connect your assets, wire them into an existing workflow, and process volume in a single pass. 100 product images with consistent treatment, 30 campaign variations, an entire asset library refresh—the workflow doesn't change, and the effort doesn't multiply. That's a different value proposition than anything Krea or Weavy currently offers.

Adobe Boards exists but doesn't do enough yet. Artlist Studio and Adobe Graph are coming. Worth watching.

ComfyUI is not an introduction. It's where you go when you've outgrown the above. Unlike the other three, it doesn't offer curated capability nodes. It exposes the actual internal components of how diffusion models work. You're not selecting "generate image"—you're assembling checkpoint loaders, CLIP encoders, samplers, and VAE decoders yourself. The ceiling disappears, and so does the floor. Get comfortable with Krea, Weavy, or Flora first. If you hit their limits and still want more, ComfyUI is waiting.

6. Notable Nodes Worth Knowing About

A few nodes across these platforms that don't exist everywhere and are worth knowing about when you find them.

Motion Transfer nodes (Krea) — capture motion from a source video and apply it to a different subject. Separate from camera control. The movement becomes the input, not the subject.

Video Time Ramp node (Krea) — remap video timing using curve types including bezier, ease-in, and ease-out. Pacing as a pipeline value rather than a post-production decision.

The LLM Call node (available across all three platforms) — puts a language model mid-workflow. Use it to analyse an image, generate a structured prompt from a brief, build a shot list, or improve a text input before it reaches a generative model. The LLM isn't just at the front of the pipeline—it can sit anywhere in it.

Bypass mode (ComfyUI) — a node state that lets you disable a node without breaking the chain. The data passes through unprocessed. The difference between Bypass and simply deleting a node is that downstream nodes still receive something to work with. Useful for A/B testing a node's contribution to a workflow without rebuilding it.

The Batch Node (Flora) — nothing else in this stack processes volume at the node level. Wire in a CSV of 100+ rows, connect your assets, run once. Everything after the idea is handled.

7. One More Thing: Adjacent Tools Worth Watching

Two things from Krea that aren't technically inside the node system yet, but are worth knowing about.

Realtime is a live generative canvas. Draw, adjust, move elements—the output updates continuously with no render queue and no button to press. Webcam, screen share, and drawing tablet are all valid inputs. It's not a node workflow, but it's the fastest possible way to test a direction before you commit to one. Sketch loose, see what the model does with it, build the pipeline once you know where you're going.

Wand extends Realtime to drawing tablet apps directly. If you're already working in a tablet environment, your marks become live generative input without leaving your tool. For anyone from a drawing or concept background, the feedback loop this creates is significant.

Neither is integrated into the node system yet. The "yet" is doing real work in that sentence.

The grammar is the same everywhere. Input, process, output. Learn it once and it transfers. The platform you start on matters less than the habit of thinking in connected steps—because once you're thinking that way, the more powerful systems stop being intimidating and start being interesting.