Diffusion Models · 5 Interactives

One Dial Between Your Image and the Model's Imagination

Text-to-image starts from pure static. img2img cheats: it pours some noise on YOUR image and denoises from there. The amount of noise is the strength dial — 0 means "change nothing", 1 means "ignore me entirely". Draw your own sketch below and find the sweet spot.

The noise dose Draw & re-dream The measured dial Two ways to fail Seeds
EP 01

Strength is literally: how much noise gets poured on your image

Before any "AI" happens, img2img runs your image forward into noise — but only partway. Strength 0.3 stops early (your image clearly survives); strength 1.0 goes all the way to static. The right panel is the actual starting point the denoiser will see. Drag the dial and watch the live correlation between your image and that starting point.

The correlation is computed live between the two panels. In real diffusion, "strength 0.6" means "jump to 60% of the way through the noising schedule and denoise from there" — fewer steps, and less of you left.

The key reframe: strength isn't a "creativity setting" bolted on top — it is literally the timestep where denoising begins. ✓ At 0.3 the denoiser inherits your composition for free. ✗ At 1.0 img2img degenerates into plain text-to-image: whatever you uploaded never mattered.
EP 02

Draw something — then let the model re-dream it

Top canvas is a pixel sketchpad — draw on it (or load a preset). Then press Run img2img: your sketch gets noised to the chosen strength and denoised step by step in the bottom canvas by a tiny "model" whose taste is bold flat shapes in this site's palette. Watch it clean up your wobbly lines at low strength, and repaint you at high strength.

Honesty label: this denoiser is a conceptual toy (blur + a palette prior), not a trained network. The mechanics of strength — noise dose, denoise from there — are exactly what Stable Diffusion's img2img does; the artwork is not.

✓ The good case in one run: at strength ≈ 0.4–0.5, your composition survives but every edge gets re-decided by the model's prior. That "keep the layout, redo the rendering" trick is why img2img is the workhorse of sketch-to-art, photo restyling, and roughing out concept art.
EP 03

Faithfulness vs creativity — put numbers on the dial

Stop eyeballing. For thirteen strengths from 0.05 to 0.95, we run the full pipeline on your current sketch and measure two things from raw pixel distance: faithfulness (how close is the output to your input) and creativity (how far did it move beyond a mere cleanup). Drag the marker; both curves are recomputed whenever you edit the sketch in EP 02.

Faithfulness = 1 − normalized RMSE(output, your input). Creativity = normalized RMSE(output, a strength≈0 cleanup of your input) — i.e. change that can't be explained as tidying. Both are live measurements, not canned numbers.

✓ Good case: the curves cross in the 0.4–0.6 band (shaded) — real transformation with your composition intact; that's why every img2img guide defaults there. ✗ Bad case: both ends of the dial — one curve is always near zero, so you're either paying for a photocopy or for a stranger's picture.
EP 04

The two ways to waste a render

Trigger both classic failures on your own sketch. Strength 0.08: the output is your input with trivially rounded corners — measure it: almost no pixels change. Strength 0.95: the "what changed" heat-map lights up everywhere and composition correlation collapses — your drawing had no say. The middle button shows the same metrics at a sane setting.

Diagnosis cheat-sheet for real tools: "it ignored my prompt and gave my photo back" → strength too low. "it ignored my photo and gave the prompt back" → strength too high. The heat-map you just used is exactly the debugging view you should imagine when tuning the real slider.
EP 05

Seeds only matter as much as strength lets them

Same sketch, same strength, three different random seeds. At low strength the three outputs are near-identical — the noise is a whisper, your image is the anchor. Crank strength up and the seeds take over: three unrelated pictures. The spread between the outputs (average pairwise pixel RMSE) is computed live as you drag.

The workflow this unlocks: pros fix the seed and sweep strength, or fix strength ≈ 0.5 and sweep seeds to harvest variations that still respect the layout. ✓ Mid strength = useful variations. ✗ High strength = the seed is choosing the picture, not you — at that point just use text-to-image.
Keep playing
Generative Models
Score Matching
Follow the arrows that point back to the data
Generative Models
Inpainting Masks
Repaint one region while the rest stays frozen