No written analysis was handed in. This summary comes from the review slides.
The question
Can one LoRA carry a style and another a form, and can mixing them render a real design? Yewon tested it on a product of their own: a planter with a growing light, rendered in the soft, white light of SANAA's architecture.
What they did
- Trained a style LoRA on photographs of SANAA buildings and a form LoRA on renders of Yewon's own Rhino model of the S-shaped planter lamp.
- Compared a text-only prompt with prompts that used the LoRAs, then mapped a grid of strengths: style from 1.2 to 2.0 with form fixed, and the reverse.
- Tested five ControlNets (DepthAnything, LineArt, Canny, HED, Binary) and settled on depth with style 1.6 and form 1.0.
- Made the final series by keeping everything fixed and changing only the noise seed.
What the tests looked at
- How much of the lamp's real shape survives when the style LoRA is turned up.
- Which kind of ControlNet keeps the form while letting the style change the material and light.
- How much one fixed setup varies from seed to seed.




































