Skip to content

Prism โ€” Vision Models

Small, self-contained image-to-image models for upscaling, denoising, and steganography โ€” general-purpose utilities that run on modest hardware.

pip install olaverse[vision]

Each Prism model ships its own small model.py architecture file alongside the checkpoint on its Hugging Face repo. Loading a Prism model downloads and executes that model.py from the corresponding olaverse/prism-* repo. All Prism repos are published by Olaverse under Apache-2.0.

PrismUpscaler โ€” Super-Resolution

Model Cards: olaverse/prism-upscaler-2x ยท olaverse/prism-upscaler-4x ยท olaverse/prism-upscaler-max

Which size?

size= Architecture Scale Pick when
"2x" (default) FSRCNN (~25K params) Fixed 2x Fast, single forward pass
"4x" FSRCNN (~25K params) Fixed 4x Bigger jump, accepts smoothing tradeoff
"max" LIIF (RRDB + implicit MLP) Any continuous resolution Exact target size needed 
from olaverse import PrismUpscaler

# Fixed scale
upscaler = PrismUpscaler(size="2x")
upscaler.upscale("input.jpg").save("output.jpg")

# Arbitrary target resolution
upscaler_max = PrismUpscaler(size="max")
upscaler_max.upscale("input.jpg", target_size=(1024, 1024)).save("output.jpg")

All three were trained with realistic degradation (blur, sensor noise, JPEG re-compression) rather than plain bicubic downsampling โ€” built for real-world low-quality input.

Known limitations

  • 4x over-smooths fine/curly hair and other high-frequency texture โ€” a consistent tradeoff at this scale.
  • None of the three have been evaluated against standard academic benchmarks (Set5/Set14/BSD100/Urban100) โ€” comparisons on each model card are informal checks against a bicubic baseline.

PrismDenoiser โ€” Noise/Blur/Compression Removal

Model Card: olaverse/prism-denoiser

Removes Gaussian noise, blur, and JPEG-like compression artifacts using a compact U-Net.

Output is always 128x128

Input is resized to 128x128 internally and returned at that resolution โ€” a 640x480 photo comes back 128x128, not restored in place. Tile larger images yourself, or chain PrismUpscaler(size="max") afterwards to reach the target resolution.

from olaverse import PrismDenoiser

denoiser = PrismDenoiser()
denoiser.denoise("noisy.jpg").save("denoised.jpg")

Reduces, doesn't eliminate, noise

On complex scenes, denoising achieves +3-4 dB PSNR but is typically incomplete โ€” some residual grain remains. On near-grayscale images, the model can render a faint color tint, since it was trained predominantly on full-color photos.

PrismSteganography โ€” Hide/Recover Messages

Model Card: olaverse/prism-steganography

Hides a recoverable message (up to 8 ASCII characters / 64 bits) inside a cover image imperceptibly, using a jointly-trained U-Net encoder / CNN decoder pair.

Save as PNG โ€” JPEG destroys the message

The hidden bits do not survive a real JPEG round-trip at any quality setting, including quality=100, and do not survive rescaling.

from olaverse import PrismSteganography

steg = PrismSteganography()

stego_image = steg.hide("cover.jpg", "hi there")
stego_image.save("stego.png")   # PNG โ€” a .jpg save loses the message

steg.reveal(stego_image)   # โ†’ 'hi there'

Images are resized to 128x128 internally; longer messages are silently truncated to 8 characters (truncation is over UTF-8 bytes, so non-ASCII input can be cut mid-character).

Measured robustness โ€” lossless only

Recovery is exact in memory, through a PNG round-trip, and under mild additive noise (Gaussian ฯƒ=5). It falls to chance under JPEG (0.42-0.48 bit accuracy across qualities 50-100) and under downscale-and-back (0.50). Treat this as a lossless-channel watermark, not a distortion-robust one. No error-correction coding is applied โ€” applications needing near-100% reliability should add redundancy (e.g. a repetition or Hamming code) on top of the raw bit channel.

Applications

  • โœ… Photo restoration โ€” upscale and denoise legacy or low-quality archives (denoise per 128x128 tile)
  • โœ… OCR preprocessing โ€” clean small scan crops before text extraction
  • โœ… Thumbnails to full-size โ€” arbitrary-resolution upscaling with max
  • โœ… Watermarking research โ€” recoverable invisible tags via steganography, within a lossless pipeline

API Reference

Full class reference: Vision โ†’