Artificial Lexicon
Tool notes

Using the walls in Stable Diffusion

Stable Diffusion gives you the dials Midjourney hides. The walls supply the words. This page shows how to weight them, negate them, and pick a checkpoint that knows them.

Weighting wall words

In A1111/Forge-style UIs, parentheses raise attention and brackets lower it: (gouache:1.3) pushes the medium harder, [background details:0.7] quiets them. ComfyUI uses the same (word:weight) syntax in its CLIP text nodes.

weighted wall wordsa fox crossing a frozen river, (gouache:1.3), (blue hour:1.15), wind-whipped snow, low-angle shot

Copy the tray output from any wall, then weight the one or two words the model keeps ignoring. Past ~1.4 most checkpoints start distorting instead of obeying.

Negative prompts

Stable Diffusion's negative box is where half the quality lives. Style words from the walls work there too. Put photorealistic in the negative when you want flat illustration, or watercolor when edges keep going soft.

💡 Keep a standing negative list per checkpoint. The walls are for the positive box; negatives are checkpoint-specific housekeeping.

Checkpoints change the vocabulary

Base SDXL knows the movement words. Anime checkpoints may not know fauvism but respond to medium words like screen print. When a wall word does nothing, suspect the checkpoint before the word. Test it on the base model.

Running locally

This site assumes nothing about your setup. The walls are plain text and paste anywhere: AUTOMATIC1111, Forge, ComfyUI, Invoke, Draw Things, or a hosted service. If you're starting out, pick whichever UI your hardware guide recommends and come back for words once it boots.

Related reading: Prompt Engineering, Properly and the Textures & Materials wall.