prep-listing-photo

Prep listing photos

Turn the user's product photos (usually clothing laid flat on a floor or rug) into clean eBay-ready images: background removed, pure white behind, a subtle drop shadow under the item, item centred on a 2000x2000 square JPEG.

Steps

1. Find the photos. Attached images live in /mnt/user-data/uploads/ (read-only). Work in the scratchpad directory, never on the originals.

2. Install the cutout tool if missing:

   pip install --break-system-packages -q "rembg[cpu]"
   

   Use the isnet-general-use model. Do NOT use birefnet-general — it is ~1GB and gets killed for lack of memory in this workspace. The model downloads from GitHub on first use, which is allowed.

3. Decide orientation first. Look at each original photo. If a shot is upside down (hood/collar at the bottom, labels inverted), list it in ROTATE below so it is turned 180 degrees BEFORE the shadow is drawn — rotating the finished image would put the shadow above the item. Tell the user which shots you turned.

4. Make the final images with this script (pass the output folder, then the photo paths). Set ROTATE to the basenames (without extension) of any upside-down shots, and SHADOW = False if the user asks for no shadow:

   import sys, os
   from PIL import Image, ImageOps, ImageFilter
   from rembg import remove, new_session
   
   ROTATE = set()        # e.g. {'IMG_2298'}
   SHADOW = True
   S, pad = 2000, 0.06   # canvas size, margin on each side
   out = sys.argv[1]; os.makedirs(out, exist_ok=True)
   session = new_session('isnet-general-use')
   for path in sys.argv[2:]:
       name = os.path.splitext(os.path.basename(path))[0]
       im = ImageOps.exif_transpose(Image.open(path)).convert('RGB')
       m = remove(im, session=session, post_process_mask=True).split()[-1]
       m = m.point(lambda v: 0 if v < 20 else v)          # drop faint specks
       if name in ROTATE:
           im, m = im.rotate(180), m.rotate(180)
       bbox = m.point(lambda v: 255 if v > 128 else 0).getbbox()
       im, m = im.crop(bbox), m.crop(bbox)
       w, h = im.size; sc = S * (1 - 2 * pad) / max(w, h)
       nw, nh = round(w * sc), round(h * sc)
       im = im.resize((nw, nh), Image.LANCZOS); m = m.resize((nw, nh), Image.LANCZOS)
       x, y = (S - nw) // 2, (S - nh) // 2
       c = Image.new('RGB', (S, S), 'white')
       if SHADOW:
           # subtle drop shadow: blurred silhouette, nudged down and slightly right
           off, blur, opacity = int(S * 0.012), int(S * 0.018), 0.28
           sh = Image.new('L', (S, S), 0)
           sh.paste(m, (x + off // 2, y + off))
           sh = sh.filter(ImageFilter.GaussianBlur(blur)).point(lambda v: int(v * opacity))
           c.paste(Image.new('RGB', (S, S), (0, 0, 0)), (0, 0), sh)
       c.paste(im, (x, y), m)
       c.save(os.path.join(out, f'{name}_white.jpg'), quality=92)
       print(name, 'done')
   

5. Check the results visually. Build a small side-by-side contact sheet of the outputs on a grey backing and look at it with Read. Check for:

   • leftover background (rug, floorboards, chair wheels, feet/slippers at the edges)
   • bits of the item cut away (hood, sleeves, drawcords, labels)
   • the shadow sitting just below the item, soft and faint, with the canvas corners still pure white (255,255,255)
   • every shot the right way up Fix anything wrong before delivering.

6. Name the files by view where it's obvious (e.g. <item>_1_front.jpg, <item>_2_inside.jpg, <item>_3_back.jpg) so they upload in a sensible order — front first, as that becomes the listing's main photo.

7. Deliver with SendUserFile (display: render), all files in one call. Finish with a couple of sentences: they're 2000x2000 on pure white with a subtle shadow, and mention any rotation or fix you made. No step recap.

Defaults and adjustments

• Pure white (#FFFFFF) background, square 2000x2000, 6% margin, JPEG quality 92, subtle shadow on.
• Stronger or softer shadow: raise or lower opacity (0.28 default; 0.4 is noticeable, 0.15 is barely there) and blur.
• No shadow: SHADOW = False.
• Different size, margin or background colour: change S, pad, or the canvas colour.
• PNG with transparency: save the cutout as PNG with the mask as alpha instead of pasting on white (no shadow).