Projects
Amusement Apparatus
Light Spring
AI Experience
Contact
Cold Maceration
AI Experience
Drawn Sept. 2026
Background replacement with a LoRA, keeping the face and the logo, wired in knob and tube
A finder and segmenter cut masks: keep the face and the logo, regenerate everything else. The operator, with a key, selects one of three preset background pictures; it and a blank card go through an image-prompt encoder. There is no text prompt. The input picture is encoded to a latent. In a loop of thirty steps, starting from pure noise, a frozen U-Net with small low-rank adapters on its attention layers predicts noise, guided by comparing the runs with the preset and with the blank; the scheduler takes one step; and the keep regions are re-injected from the original latent, re-noised to the new step. After the last step the decoder makes pixels, the original face and logo pixels are pasted back, feathered outward, and a check confirms they are identical to the input. An inset shows the adapter trained beforehand on photos of the three backgrounds, in its own loop, the gradient updating only the adapter.
DENOISING LOOP · 30 STEPS
LoRA TRAINING · BEFOREHAND
INPUT
OUTPUT
FACE & LOGO FINDER · SEGMENTER
KEEP
REGENERATE
1
2
3
BACKGROUND · ONE OF THREE PRESETS
BLANK
OPERATOR’S KEY
PRESET ENCODER
IP-ADAPTER
NOISE · SEED 42
IMAGE ENCODER
VAE
LATENT z₀
DOWN
DOWN
DOWN
MIDDLE
UP
UP
UP
PRESET → CROSS-ATTENTION · TEXT PROMPT LEFT EMPTY
U-NET · FROZEN W · LoRA A·B ON ITS ATTENTION LAYERS
GUIDANCE × 7
WITH PRESET / BLANK
SCHEDULER · t
RE-NOISE z₀
TO STEP t−1
STEP
z(t) → z(t−1)
RE-INJECT
KEEP REGIONS
↓8
× 30
FIRST PASS: PURE NOISE
IMAGE DECODER
VAE
PASTE BACK
ORIGINAL FACE & LOGO
FEATHERED OUTWARD
PIXEL CHECK · IDENTICAL TO INPUT
PHOTOS OF THE 3 BACKGROUNDS
+ NOISE
FROZEN W + A·B
LOSS
NOISE ERROR
GRADIENT → A·B ONLY
× 2,000 STEPS
A·B
coldmaceration