28
PRACTICE LAB · GENERATIVE SYSTEMS
Generative Bench — one frame, one seed, one variable
This lab closes Global Prompts, SDXL, ControlNet, IPAdapter/LoRA, masks and detail conservation. The key skill is not “turn everything at once,” but running controlled A/B tests in which exactly one cause changes.
CONFIRMEDARCHVIZ FOUNDATION / EXTENSION. Specific Hansen settings are used as a reference baseline; the learning goal is a transferable experimental method.
One benchmark — one changed variable
Prompt: separate content from style/light
- Lock seed and geometry controls.
- Create a baseline with the current GLOBAL prompt.
- Change only LIGHT STYLE or one material descriptor.
- Compare not “better/worse,” but what actually changed: material, light, composition or geometry.
SDXL: prove the role of latent source and denoise
- Create a Mode 1 baseline.
- Switch to the image-seeded variant with the same seed and minimal denoise.
- Increase denoise in steps without changing prompt/control.
- Record the threshold at which architecture starts to drift.
Depth vs Canny: they preserve different kinds of structure
| Run | Control | What to evaluate |
|---|---|---|
| A | Depth only | Volume, perspective, major geometry |
| B | Canny only | Contours, frames, fine facade lines |
| C | Depth + Canny | Trade-off between structure lock and freedom |
IPAdapter / LoRA: enable only after the baseline is proven
- Prove the baseline without reference conditioning first.
- Add the reference and begin with a low weight.
- Compare whether the intended visual language transfers without damaging architecture.
- If the result cannot be explained by one variable, the test is invalid.
Prove the mask as data before using it for an edit
- Generate the mask and expose it in a dedicated preview.
- Check boundaries before connecting inpaint/composite.
- Run grow/blur A/B and evaluate only edge quality.
- Only then use the mask in the downstream operation.
Detail transfer: find the minimum sufficient strength
- Compare 0 / baseline / increased strength with the same seed.
- Inspect windows, facade rhythm, joints and small structural elements.
- Choose the lowest value that genuinely restores detail.
LAB 05 is complete when every result can be explained causally
- Every comparison has a baseline.
- Only one controlled variable changes in each run.
- Seed and input are fixed.
- Intermediate previews exist; evaluation is not limited to the final frame.
- You can identify which module caused a specific change.