Input data — why a production workflow does not begin with the prompt
In this part of the showcase, Hansen presents prepared input data before generation. For ArchViz, this is fundamental: the workflow receives not only text, but a set of visual sources carrying geometry, structure, masks, references and downstream controls.
What is presented as input data
The video shows several types of prepared visual information: a stylized/false-color render, a depth-like pass, an RGB-coded mask-like pass, reference/final-look images and the 3D scene context itself. This demonstrates a production-first approach: AI does not have to infer everything from a single prompt.
The source graph confirms several independent LoadImage sources: BASE IMAGE, external depth, IPAdapter references, RGB-coded mask source and final overlay assets.
INPUT CONTRACT: every file should have one clear role
Mandatory base
Source notes identify node 79 as BASE IMAGE by connectivity and an embedded workflow note.
Optional means branch-dependent
Other inputs are optional only when their consuming branch is bypassed or selected away; “optional” is a routing property, not an intrinsic property of the file.
| Input type | Source evidence | Role |
|---|---|---|
| BASE IMAGE | node 79 | Primary architectural source; feeds resize, preprocess, compare and mask paths |
| External depth | node 25 | Alternative geometry/depth source selected downstream |
| Reference images | nodes 41 / 42 | Optional IPAdapter reference inputs |
| RGB-coded masks | node 338 | Source for mask extraction by color |
| Logo / overlays | nodes 301 / 309 | Final delivery overlay stage |
Not every image in a workflow means the same thing
For a beginner, one of the most important habits is to look past the thumbnail and ask what information the image carries and which downstream node consumes it.
| Visual input | Human interpretation | Machine role |
|---|---|---|
| Beauty / base render | What the scene looks like | IMAGE source |
| Depth | What is nearer / farther | Geometry guidance |
| RGB ID / mask pass | Which region belongs to what | Region selection |
| Reference image | Desired visual language | Style / appearance guidance |
| Logo / overlay | What to add to delivery | Post-process asset |
Dimensions and coordinate space are part of the input contract
Images that look identical can be incompatible when they have different dimensions, aspect ratios or coordinate spaces. Resize 783, external-map sizes and later detection canvases are therefore part of the architecture, not implementation trivia.
Exercise: inventory the inputs
- Find every LoadImage node in the INPUTS area.
- For each one, record its function rather than its filename: BASE / DEPTH / REFERENCE / MASK / OVERLAY.
- Trace only the first downstream link from each input.
- Mark which inputs participate in the active route and which belong to bypassed/optional branches.
When the 01:08 lesson is complete
You understand that production ComfyUI begins with input-data design. A prompt is only one input to the system, not the entire system.