Compression using Implicit Neural Representation

How does it work?

A tiny neural network learns to memorize a single image — storing everything needed to repaint it, pixel by pixel, from scratch.

Embeddings A low-resolution grid of learned tiles — one per region. Each tile encodes what this area looks like: a blurry hint the network looks up by position. A grid of learned region hints.

Neurons Three small layers of arithmetic that read the tiles and learn how to paint them into real R·G·B pixels. Each layer refines the result a little more. Layers that paint pixels from hints.

Training

Construct and compare.
Score the error.
Nudge numbers.
Repeat thousands of times.

The image gradually sharpens.

Fun things to try

Pause training and click a single embedding tile — see its isolated influence on the image.
Switch the activation to None and watch the network struggle without non-linearities.
Paint a region with the ROI Mask to tell the network what to preserve in priority.
Hit Shake occasionally — it jolts the model out of a local minimum.

click · any key · to continue

Some features are reduced on mobile

Source

⊕

Drop Image
or Click

Input Size —

Layers update every steps

Loss — Step 0 Rate —

Output

1×

Compressed Size —

grid=N	Embedding grid resolution — 16, 32, or 64
EMB=N	Embedding channels per cell — 4, 8, or 16
MLP1=N	Layer 1 output width — 4, 8, 16, 32, or 64
MLP2=N	Layer 2 output width — 4, 8, 16, 32, or 64
iters=N	Max training steps (0 = run forever)
embedLr=N	Embedding learning rate slider value (0–100)
mlpLr=N	MLP learning rate slider value (0–100)
smooth	Enable smooth interpolation
nosmooth	Disable smooth interpolation
8qat	Enable 8-bit quantization-aware training for the MLP
4b	Pack embeddings as 4-bit (requires EMB ≥ 8)
8b	Pack embeddings as 8-bit (default)
none	Disable MLP quantization (fp32 weights)
numLoops=N	Number of times each EMB/MLP phase is repeated per super-cycle (1–8)

Embedding Grid	Spatial resolution of the feature grid (N×N). Larger = finer detail, bigger model.
Embedding Channels	Feature dimensions per grid cell. More channels = richer representation.
MLP Width 1	Output width of Layer 1 (embCh → W1). Larger = more capacity, slower training.
MLP Width 2	Output width of Layer 2 (W1 → W2). Independent from W1.
Activation	Hidden-layer activation function. SIREN (sin) works well for natural images. Tanh and Softsign are smooth alternatives. None (identity) is useful for educational purposes — makes the MLP a purely linear model.
Quantization	MLP weight precision. QAT trains with simulated 8-bit quantization, yielding a much smaller GLSL export.
Emb Bits	Quantization depth for the embedding grid. 4-bit halves storage vs 8-bit.
Smooth Interp	Bicubic-like grid sampling via IQ's smooth-step trick. Improves quality at negligible cost.
No Offset	Disable per-plane UV jitter. Each embedding plane normally samples at a small random offset to increase effective resolution.

Max Iterations	Training stops after N steps. Set to 0 to run indefinitely.
Embeddings LR	Adam learning rate for the embedding grid (log scale: 1e-4 … ~1.2e-1; default 3.0e-2).
MLP LR	Adam learning rate for MLP weights (log scale: 1e-4 … ~1.2e-1; default 3.0e-2).
Loss	MSE over RGBA with ITU-R BT.601 luminance weights on R/G/B (0.299 / 0.587 / 0.114). Alpha is weighted equally and supports transparent images.
Emb / MLP ratio	How many embedding-only Adam steps are taken per MLP step. Higher = focus training on the grid.
Num Loops	Repeats each phase N times per super-cycle. With ratio=5 and loops=3: 15 consecutive EMB steps then 3 MLP steps, then repeat.
Bwd pixel stride	Subsample every Nth pixel during backprop. Speeds up steps at the cost of noisier gradients.

Brush Radius	Size of the paint brush in pixels.
ROI Strength	Loss multiplier inside painted regions. Higher = stronger bias toward those areas.
Freeze Mask	Prevent the mask from decaying over training steps.

COIN	Dupont et al., 2021 — COIN: COmpression with Implicit Neural representations
Instant-NGP	Müller et al., 2022 — Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
EPFL CVLab	EPFL CVLab — Image Compression using Implicit Neural Representations
Kodak set	Kodak lossless true color image suite — predefined example images

Source	Input image. Click or drop a new image to replace it. Paint the ROI mask by dragging.
Embeddings	All embedding channels shown as grayscale thumbnails. Click a thumbnail to flip all other channels in the forward pass, leaving the clicked one unchanged.
Layers	Inference flow diagram: embedding → L1 → L2 → L3 → output. Pre-activation values are shown per-neuron with per-layer min/max normalisation. Updates every few steps.
Output	Decoded result at full resolution. Use the zoom slider to upscale.

↑ Load	Load a .safetensors model file (also drag-and-drop onto the page). Restores weights and all UI settings.
↓ Save	Save model weights and current UI settings as .safetensors.
↓ GLSL	Export the trained model as a self-contained GLSL fragment shader (ShaderToy-compatible).
● Snapshot	Store current weights as an in-memory checkpoint.
↺ Recall	Restore weights from the last snapshot (enabled only when a snapshot exists).
⚡ Shake Emb	Add small noise to the embedding grid to escape local minima. Resets embedding Adam moments.
⚡ Shake MLP	Add small noise to all MLP weights and biases to escape local minima. Resets MLP Adam moments.
↺ Reset Model	Reinitialize the model and optimizer from scratch (Training panel in sidebar).
Space	Toggle training on/off (Stop / Continue).