fluxpediawork/blob-portrait

Blob portrait: gradient descent for a photo

Paints a photo with 2D Gaussian blobs. It starts from 1,500 random blobs, and each step it renders them, measures the error against the photo (L1 + SSIM), and moves every blob a little to reduce it (Adam). New tiny blobs spawn where the error is highest. No image generator is involved: the picture is learned.

Steps

1. Get the photo. Use the attached file. Note its subject position so you can set --focus (the area that gets the most close-up training, usually the face).

2. Install. Python 3.10+.

   pip install -r requirements.txt   # torch, numpy, pillow
   

   If the PyTorch index is blocked, plain pip install torch from PyPI works.

3. Pick a preset from the hardware and time available:

   preset	steps	max blobs	2-core CPU	GPU
   quick	300	15,000	~8 min	~1 min
   default	1,600	45,000	~35 min	~5 min
   full	6,000	100,000	hours	~20-40 min

   On a CPU, run default in the background and poll the log; don't block on it.

4. Run.

   python scripts/blobfit.py photo.jpg --out result/ --preset default \
       --aspect 4:5 --focus 0.5,0.4,0.45,0.5
   

   • --aspect crops to w:h around the centre (none keeps the photo's own shape).
   • --focus cx,cy,w,h is in 0-1 of the cropped image. For a centred selfie, 0.5,0.4,0.45,0.5 covers the face.
   • --steps / --blobs override the preset.
   • Device is picked automatically: CUDA, then Apple MPS, then CPU.

5. Check the result before delivering. Open compare.jpg and zoom.jpg and look at them. stats.json has the PSNR: around 28-30 dB is a soft likeness (quick), 38 dB+ is hard to tell apart at normal size. If the face is soft, rerun with more steps or move --focus onto the face.

6. Deliver. Send compare.jpg, render.png and training.gif. Say how many blobs and steps, the PSNR, and that it was trained, not generated.

Outputs

file	what it is
render.png	the blob painting at full fitted size
compare.jpg	real photo beside the code version
zoom.jpg	close-up proof around the eyes
training.gif	noise to picture, one frame per milestone
blobs.npz	every blob: position, scale, angle, colour, strength (for re-rendering or animation)
stats.json	steps, blob count, PSNR, seconds, device

Honest limits

• It copies one photo. It can't invent a new view or a new face.
• Quality depends on steps: short CPU runs come out soft; fine skin grain needs the full preset.
• The backdrop colour is taken from the photo's edges, so blobs only need to paint what differs from it.