Experiment 04 — LoRA Style Wedge

Prompts

Jon → Hermes (spoken input)

"One text prompt, run it four times. Each time use a different LoRA. The LoRA name is the trigger word — so warm-miniature-diorama is the trigger. Pick a simple prompt like a cat sitting on a kitchen counter, render four times with four different looks."

Hermes → ComfyUI (KREA2 turbo t2i — 4 runs)

[trigger word], a cat sitting on a kitchen counter, morning sunlight through the window, realistic kitchen with granite countertop, tiled backsplash, a coffee mug nearby Run 1: warm-miniature-diorama, ... Run 2: surreal-photorealistic-composite, ... Run 3: misty-dark-fantasy-render, ... Run 4: iridescent-crystal-bokeh, ... Seed: 42 (same for all 4) Model: KREA2 turbo fp8 Chain: UNETLoader → TextFusion LoRA → Style LoRA → KSampler Steps: 8 · CFG: 1 · Prompt refinement: disabled

Results — 4 LoRA Styles

Warm miniature diorama style — LoRA: warm-miniature-diorama

Surreal photorealistic composite style — LoRA: surreal-photorealistic-composite

Misty dark fantasy render style — LoRA: misty-dark-fantasy-render

Iridescent crystal bokeh style — LoRA: iridescent-crystal-bokeh

tldraw Canvas — GUI Export

tldraw canvas04 GUI export showing LoRA wedge — Canvas 04 exported from tldraw GUI — 2×2 grid of LoRA variants with prompt text and description, showing the tldraw desktop app workspace during the experiment.

What Happened

Setup: Four KREA2 style LoRAs, each 224 MB, installed in models/loras/krea2/. The workflow chained two LoRA loaders: the existing TextFusion Refusal Reduction LoRA (for prompt compliance) followed by the style LoRA, both feeding into the KSampler.

The base prompt was identical across all four runs: "a cat sitting on a kitchen counter, morning sunlight through the window, realistic kitchen with granite countertop, tiled backsplash, a coffee mug nearby." Same seed (42). The only variables were the LoRA file and the trigger word prepended to the prompt.

Jon caught an inaccuracy: The canvas description said "only the LoRA changes" — but that's not quite right. Both the LoRA file and the trigger word in the prompt change together. The trigger word activates the LoRA's learned style; without it, the LoRA weights apply but the style doesn't fully activate. So saying "only the LoRA changes" is misleading — the prompt also changes by one word.

Generation: All 4 images generated in 33 seconds total (~8s each, models warm). Vision analysis confirmed distinct styles:

warm-miniature-diorama: Classic tilt-shift / miniature faking effect — extremely shallow depth of field, warm golden lighting, the kitchen looks like a tiny toy model.

misty-dark-fantasy-render: Moody, foggy, volumetric god rays streaming through the window, desaturated palette, dark fantasy atmosphere — like a concept art still from an indie game.

surreal-photorealistic-composite and iridescent-crystal-bokeh each applied their own distinct visual treatment to the same scene.

Labels simplified per Jon's request: Initially each image had both the LoRA name and the trigger word. Jon said to remove the trigger — just show the LoRA name. The labels were updated in-place via the tldraw API.