EP4 · product photography · free episode kit
I Asked Nano Banana for a White Background — It Redesigned My Logo
A white background should not mean a new logo. This kit takes a different route: let AI estimate transparency, then place the original product pixels on white.
Watch EP4
The video link will appear here after EP4 is published. The free kit and notebook are ready below.
The three-step workflow
- Make a mask, keep the photograph. BiRefNet-Lite estimates which pixels belong to the product. Its output is an alpha mask, not a newly painted shoe, label or logo.
- Manage color and put it on white. Apply EXIF orientation, convert the embedded ICC profile to sRGB, then composite the source RGB at original resolution. If the source has no ICC profile, stop and confirm its color space before opting into the sRGB assumption.
- Check before you deliver. Read the machine report, then compare logo, lettering, mesh, laces, holes and edges against the original at 4×. A passing pixel check cannot tell you whether the mask selected the right edge.
Get the exact workflow
The free Lite edition is the episode kit: one photo, a white background and checks. Full adds conveniences for repeated work. Both editions use the same pinned segmentation model and preserve the default opaque product core after color conversion.
Start here
Free Lite + notebook
$0
- Process one photo at original resolution.
- Automatic mask, ICC color management and pixel checks.
- Local CPU or compatible NVIDIA GPU.
- Optional free notebook in your own cloud account.
Download the Lite ZIP first, then follow the README or upload the ZIP and your photo in Colab. No payment required. Free cloud capacity is not guaranteed.
For repeated work
Full package
$9.99 USD · one time
- Everything in Lite.
- Batch manifests and canvas presets.
- Optional programmatic contact shadows.
- Optional tone lift with measured color checks.
Full ZIP delivery through the Whop Files tab after checkout. Download access: 120 days from purchase. Keep your own copy. No continuing updates promised.
Run locally, or choose your own cloud
Local CPU is the default. A compatible NVIDIA setup automatically uses CUDA; initialization errors stop instead of silently switching devices. Photos stay on your computer during local processing; the first run downloads the fixed model, approximately 224 MB.
The free notebook uses your own Colab account. Download the Lite ZIP from the public repository, open the notebook, then supply the ZIP and one authorized photo. Choosing cloud uploads your inputs to that account. Keep them private. Free quotas are not guaranteed.
What the evidence covers
The underlying 0.2.0 scripts passed eight Linux CPU/CUDA runs across Full and Lite on two 4032 × 3024 photographs. The 1.0.0 release keeps those processing scripts and dependencies unchanged. CPU whole-command time in that test was 23.82–26.37 seconds per photo, with peak process-tree memory of 6.57–6.72 GiB. These are test observations, not a speed promise.
Windows clean installation and the notebook’s interactive Colab upload/download flow have not been fully tested. Start testing with 16 GB RAM; this is not a proven minimum, and an 8 GB system may run out of memory.
The title describes this episode’s comparison; it is not a claim that every generative image tool always fails. No perfect masks, marketplace approval or zero color difference are promised. Optional Full tone lift deliberately changes color values.