A 22-minute ComfyUI video turns a gray SketchUp viewport into a warm residential exterior. The creator drops in a workflow file, mentions ControlNet, changes three values, and gets a convincing result. You download the graph. Six nodes turn red. The checkpoint is not named. The depth input in the demonstration is not included. The output you get has a new balcony and one fewer window.
Today's community sweep surfaced a familiar request across architecture and ComfyUI forums: people want a workflow that adds lighting, materials, vegetation, and reflection detail without losing the design. There are plenty of impressive demonstrations. There are far fewer packages another person can run, diagnose, and adapt.
That difference is reproducibility. For architectural work it is not academic housekeeping. It is the line between a tutorial and a magic trick.
The image proves less than the package
A final image proves that one output existed on one machine. It does not prove which graph made it, whether the creator selected the best of eighty attempts, or how much changed in Photoshop afterward. Architecture adds a stricter requirement: the output must preserve specific openings, edges, proportions, and relationships. A beautiful image can still be a failed architectural result.
Judge a tutorial by the material needed to reproduce one representative pass. That package should allow a second user to load the graph, resolve its dependencies, feed it the same source, run it with the same settings, and understand why the result differs if it does.
The workflow file is not the workflow. The workflow is the graph plus every dependency and decision required to make it behave.
The seven-item reproducibility checklist
1. An exported graph, not a screenshot
The tutorial must provide the ComfyUI workflow JSON or a PNG that contains valid workflow metadata. A screenshot is useful for orientation but cannot preserve node values, links, widget states, or hidden groups. Load the supplied file into a clean ComfyUI session before watching the video. Missing nodes should become visible immediately.
The graph should use labeled groups such as Input, Structure Control, Generation, Refinement, and Export. Labels are not decoration. They show where to replace an input or disable a stage without tracing cables across the canvas.
2. Exact model identifiers
"Use FLUX" is not enough. Record the checkpoint filename, model family, variant, precision, VAE, text encoders, ControlNet model, upscaler, and any LoRA files. Include the source page or repository and a checksum when practical. Two similarly named files can produce different results or require different loaders.
Licensing belongs in the same list. A tutorial that works only with a model your practice cannot use commercially is a learning exercise, not a production recipe.
3. A custom-node manifest
List each custom-node repository with the installed commit or release. "Install missing nodes with Manager" can get a personal workstation running, but it does not freeze a studio setup. A repository can change interfaces, rename a node, or pull a dependency version that breaks another graph.
Test the manifest in a separate ComfyUI installation or virtual environment. If the graph loads only on the creator's long-lived machine, undocumented residue is doing part of the work.
4. Every source input
Share the base render or viewport, masks, depth map, line pass, reference image, and any alpha channel used in the demonstration. If project confidentiality prevents sharing the original, provide a small substitute scene that exercises the same controls. It should include thin frames, repeated windows, glazing, planting edges, and at least one difficult occlusion.
Explain how each control image was produced. A depth pass exported from the 3D model is not interchangeable with depth estimated from a flat image. The first encodes actual project geometry. The second is an inference that may misread glass, sky, or overlapping masses.
5. The values that govern structure
Record resolution, seed, sampler, scheduler, steps, CFG or guidance, denoise strength, ControlNet strength, start and end percentages, resize behavior, and mask feathering. For architecture, denoise and structural-control values deserve special emphasis because they determine how much permission the model has to redesign the building.
Do not present a single value as universal. Show a narrow working range and what fails at either edge. If denoise at 0.25 preserves the facade but looks flat, while 0.55 invents windows, that boundary teaches more than the chosen 0.38.
6. Hardware and runtime
State the GPU, VRAM, system memory, operating system, ComfyUI version, image size, and measured runtime. Note any quantization, tiled VAE, attention mode, or low-memory option. A graph that runs in four minutes on 24 GB of VRAM may fall back to system memory or fail on an 8 GB laptop.
Provide a lower-memory branch when possible. Reducing resolution, using tiled processing, or disabling the final upscale can preserve the lesson even when it cannot match the demonstration's output size.
7. A geometry comparison
Put the output beside the input and mark the architectural changes. Count windows, compare roof edges, inspect stair direction, trace balcony lines, and check whether adjacent views still agree. A tutorial about enhancement should state its acceptable change threshold. "Looks similar" is too loose for project work.
| Package item | Minimum evidence | Architecture check |
|---|---|---|
| Graph | Loadable JSON or metadata PNG | Structure stage clearly labeled |
| Models | Exact files, sources, license | Control model matches input type |
| Nodes | Repositories plus versions | Clean install loads without red nodes |
| Inputs | Base image, controls, masks | Depth origin is documented |
| Settings | All generation and control values | Safe denoise range is shown |
| Runtime | Hardware, resolution, elapsed time | Lower-memory path exists |
| Validation | Input-output comparison | Geometry changes are marked |
A 30-minute tutorial audit
Before investing a day in a workflow, give it half an hour. Spend five minutes checking that the graph and inputs are downloadable. Spend ten loading dependencies in a clean session. Spend ten running the included example without changing the prompt. Spend five comparing the output against the source geometry.
If the graph cannot clear that audit, it may still contain a useful technique. Treat it as a reference, not a recipe. Rebuild the relevant section in your own known graph and document the missing choices yourself. This prevents a studio library from filling with brittle JSON files nobody trusts.
How creators can publish a better workflow
A strong release does not need a long course. One folder can contain the graph, a README, a model manifest, the sample inputs, a reference output, and a small troubleshooting section. The video can then explain judgment: why depth is stronger than edges in one zone, why denoise stops below a certain point, and which errors signal that the model is departing from the building.
Version the folder whenever a node or model changes. Keep the earlier release available. A date and version number let a viewer match the files to the demonstration instead of guessing whether today's graph is the one shown six months ago.
Our take
ComfyUI remains valuable precisely because its graph exposes the process. Yet exposed nodes do not automatically create transparency. Unnamed models, absent inputs, floating dependencies, and unreported selection can make an open graph as opaque as a one-click service.
The community does not need another perfect reveal with the wires cropped out. It needs graphs that load, inputs that travel, settings that teach boundaries, and comparisons that admit what moved. If another architect cannot rebuild it, you watched a result, not a tutorial.
Written from the 30 August 2026 intel sweep, which surfaced repeated requests for ComfyUI architectural enhancement tutorials that improve materials, light, planting, and reflections while preserving project geometry. ArchiGen AI carries no sponsored placements.