Grande Station Library or Steampunk Station

Grande Station Library or Steampunk Station An examination of SketchUp’s Diffusion extension:

SketchUp 2020, SU4iPad, SU2023 Diffusion & Thea Render:

Conclusion:
A very fast, powerful, one eyed, nearsighted thoroughbred race horse with blinders on but, wonderfully fun to play with. My findings are described below as to why that is my conclusion along with my observations of it’s strength and weakness in context of a large model.

The Project:
A collage of personal work brought together for these Diffusion tests, was originally started back in 2019 from a vision of a much larger, expansive project combining a great library, social hub and train station mixing steampunk and ancient, extinct, titanic civilizations megalithic stone structures in the background to include allot of air transport with a focus on this train. The train itself is an ongoing project that is not meant to be presented strapped to tracks, but instead, floating locomotion via electromagnetic generators. It’s inspiration comes from the work of Arcadeous Phoenix, who also built the original rail system here (now modified) and the transformers sticking out of the roof. Nothing here is completed. The train has work to be done, the station is not complete (piping, ventilation, accessories, tunnel entrance/exit points, smaller square windows frames to fit in these larger roof frames, etc.)

However, in the interest of working with a larger scale model, these original assets were edited to present this version specifically for these Diffusion tests.

Motivation:
Testing Diffusion on smaller models and seeing so many examples of Diffusion created by other users, I was curious how well the extension would handle allot of geometry in a larger scale project, principally viewed from the interior. This is the answer to those musings. For my previous comparison of Diffusion to the one eyed horse, I’ll provide the breakdown as fallows.

Fast and Powerful:
It truly is remarkable to see the variation, effects, bells & whistles, additions and vast colour schemes that the extension produces so quickly. With every click of “Generate”, even without adding new descriptions, Diffusion creates new images with a wonderful set of three variation renders.

The pros and cons with an exercise like this are to be balanced as with scales, new inspirational ideas in one hand and loss of progress on what is desired in the other. With each new generation, three new options are provided with amazing speed. Comparatively, this is why I’m using thoroughbred racehorse as a label. For this aspect, Diffusion is truly impressive for use in concept to convey a vision without the need to be too specific or faithful to detail or respecter of already chosen materials.

Yet, it is fast and powerful.

One eyed, nearsighted:
As seen from the comparison images from SketchUp and Thea Renders, there’s much detail that gets lost. When the scene’s center has geometry further off, Diffusion seems to get a bit confused as if it’s nearsightedness works great for what is directly in front of it but that distance seems to confuse the AI. So, simple objects, centered and close produce rather (or more so) reliable representations while distant objects produce scattered or blurry results. But, even with something as simple as a round, circle port window, Diffusion translates the geometry into something seen in a Salvador Dalì painting.

Even though the simple geometry is right in front and center… this is where I wish the maximum 100% value of “Respect Model Geometry” was reworked so that 100% actually is 100% and not a 50% “close your eyes, fill in the gap, hit and miss, prancing pony” thing. In this case, I really want to rely more on the horse power this beast has but, isn’t applied to the extend I think it should bed in this aspect.

Blinders:
As mentioned before about focusing on near/far objects, I find that it seems to be accentuated even more with stuff on the sides. It seems (perhaps my perspective anyway) that the AI gets even more funky, creative at filling in the gaps with the stuff that is not right in front and more to the sides. I haven’t quite figured out to what extent or if it’s just a continuation of the same issue mentioned above. But, again, this might be confirmed or solved if selecting 100% on “Respect Model Geometry” will be reworked to do just that.

One obvious discovery, when tried, is that Diffusion currently rejects 2pt perspective. After preparing scenes with 2 pt perspective to keep verticals in certain scenes, the launch of Diffusion kicks that perspective right out. So, the plugin cannot take advantage of SketchUp’s different camera features. As long as one knows this, the scenes can be set up appropriately to accommodate this current weakness.

However, I conclude that Diffusion is an excellent tool and amazing resource for concept work. To be able to produce, in shorthand and scribbles, an image that conveys “ideas” of what something “could” look like or be in t