Reference Sheet Creation for IC-LoRA Ingredients

Proven workflow for creating IC-LoRA reference sheets from stock character images on TrueNAS. Used 2026-07-22 for the boss+woman cyberpunk test.

Steps

1. Download stock from TrueNAS

smbclient -N //10.0.0.117/proxmoxBackup -c 'cd ai_vid_stock_material\\character_refs; get SHEETS2_00005_Boss.png /tmp/SHEETS2_00005_Boss.png'
smbclient -N //10.0.0.117/proxmoxBackup -c 'cd ai_vid_stock_material\\character_refs; get cyberpunk_woman_neon_01.jpg /tmp/cyberpunk_woman_neon_01.jpg'

2. Extract panels from multi-panel character sheets

The boss stock (3328×2432) is a 2-panel sheet on a bright background. Use brightness thresholding to find content regions:

from PIL import Image
import numpy as np

img = Image.open("/tmp/SHEETS2_00005_Boss.png")
arr = np.array(img)
brightness = arr.mean(axis=2)
is_content = brightness < 200  # content is darker than bright background

# Find vertical regions (rows with >5% content pixels)
row_content = is_content.mean(axis=1)
content_rows = np.where(row_content > 0.05)[0]
gaps = np.diff(content_rows)
split_points = np.where(gaps > 20)[0]  # gaps >20px = panel boundary

# Extract each panel with horizontal bounds
regions = []
start = content_rows[0]
for sp in split_points:
    end = content_rows[sp]
    regions.append((start, end))
    start = content_rows[sp + 1]
regions.append((start, content_rows[-1]))

for i, (y1, y2) in enumerate(regions):
    region = is_content[y1:y2+1, :]
    col_content = region.mean(axis=0)
    content_cols = np.where(col_content > 0.02)[0]
    x1, x2 = content_cols[0], content_cols[-1]
    crop = img.crop((x1, y1, x2, y2))
    crop.save(f"/tmp/boss_panel_{i}.png")

3. Composite reference sheet

IC-LoRA requires: 768×448, black background, one panel per character, NO text.

ref = Image.new("RGB", (768, 448), (0, 0, 0))

# Place boss panels (left side)
bp0 = boss_panels[0].copy()
bp0.thumbnail((300, 200), Image.LANCZOS)
ref.paste(bp0, (10, 10))

bp1 = boss_panels[1].copy()
bp1.thumbnail((300, 220), Image.LANCZOS)
ref.paste(bp1, (10, 220))

# Place woman panels (right side)
w1 = woman1.copy()
w1.thumbnail((200, 200), Image.LANCZOS)
ref.paste(w1, (330, 10))

w2 = woman2.copy()
w2.thumbnail((200, 200), Image.LANCZOS)
ref.paste(w2, (550, 10))

ref.save("/tmp/ic_lora_reference_sheet.png")

4. Loop to 121-frame static video

ffmpeg -y -loop 1 -i /tmp/ic_lora_reference_sheet.png \
  -c:v libx264 -t 5.04 -r 24 -pix_fmt yuv420p \
  /tmp/ic_lora_reference_121f.mp4

5. Copy to .202 for ComfyUI

sshpass -p 'passw0rd' scp /tmp/ic_lora_reference_121f.mp4 [email protected]:~/comfy-ui/input/

Pitfalls

Bright backgrounds need thresholding. The boss stock has a ~230 brightness background — content detection needs brightness < 200, not < 30 (which is for black backgrounds).
Panel gap detection is fragile. The gaps > 20 threshold works for the boss sheet but may need tuning for other sheets. Always inspect extracted panels before compositing.
IC-LoRA requires black background. The reference sheet MUST have a black background — bright backgrounds confuse the conditioning.
Bigger panels = better carry-over. Give important characters more space in the composite.
Resolution must match trained bucket. 768×448 is the IC-LoRA trained resolution. Other resolutions may work but are untested.

3.4 KiB Raw Blame History Unescape Escape