Diffusion Illusions | One Sheet, Two Pictures

Gallery

40 illusions, each scored 4 or 5 in a blind review. 14 are keepers from export window2-2026-08-clean. The rest scored as well but come from other exports and middle stages, and most have no frame rating yet. Click an image, or Flip, to turn the sheet. A good seed does not promise that the same pair will work on a new one.

What you are looking at

A flip illusion is one printable picture. Upright, it shows subject A. Turn the sheet 180 degrees. It shows subject B. The prime is the file for the printer. The two views on this page show that same sheet before and after the turn.

The method is Diffusion Illusions (Burgert et al.). A frozen image model scores how each view matches its prompt. Only one small network trains. It paints the prime. The large model stays frozen.

This page is the public study of that method on our worker stack. A typeset paper can follow. You can read this version now.

**Figure 1.** Where this method sits. Sampling-time methods combine noise estimates inside one diffusion pass. This page optimises a prime network instead, which costs far more and accepts arrangements that are not pixel permutations. The gallery at the foot is Figure 1 of Illusion3D (Feng et al., arXiv 2412.09625), cropped, used under CC BY 4.0.

How the optimizer is built

The optimizer has four parts. A prime network paints each printable image. Fixed arrangements turn primes into derived views. A frozen diffusion model scores those views. Only the prime network trains. In code, the parts have these names: FourierFeatureNetwork, the ILLUSIONS flip arrangement, and DiffusionAdapter. The paper calls them prime images, arrangement processes, and the frozen model.

Flip uses one prime and a 180-degree turn. Rotate and hidden types stack transparencies. This page shows flip only. We measured the flip type.

Gradients never flow through the UNet. The large model stays frozen. First comes Score Distillation. Then comes Dream Target.

**Figure 2.** One prime, two arrangements, two views. Frozen SD 1.5 scores the views. Only the prime weights move.

**Figure 3.** How we bake one cell. Make primes. Make flip views. Run Score Distillation. Reset Adam. Run Dream Target. Out come the prime and both views.

The prime is a small network

Each prime is a Fourier Feature Network. Pixel coordinates go through a fixed random projection, then sine and cosine, then a small network, then sigmoid RGB. The default network uses 256 Fourier features at scale 10. Its MLP is 512-256-256-256-3, which is 263,683 weights, and it renders at 512 px. Every image on this page came from the research network instead. That one uses 128 features, four 1x1 convolutions with BatchNorm, 199,683 weights, and it renders at 256 px. Only these weights move.

The paper shows why this matters. Pixel optimization hides the signal in high frequencies. A printer cannot hold that noise. Network-weight optimization keeps the signal in printable structure.

**Figure 4.** Pixel coordinates go in. Printable RGB comes out. The chain shows the network that painted this gallery. The CLI default differs, and the figure says how.

Score Distillation

Phase 1 uses frozen Stable Diffusion 1.5. The step encodes each derived view to latents. It adds noise at a random timestep. The frozen UNet scores the noised latents with high classifier-free guidance. One doubled UNet forward covers all views, unless the view batch size splits it into chunks. The default guidance is 100. This gallery ran 60. Timesteps are integers between 2 percent and 98 percent of the schedule.

The step applies the guided noise residual as a gradient on those latents. That gradient updates only the prime network. The paper printed pseudocode has no gradient path. The authors' code uses residual-as-gradient. Every SDS implementation uses that form.

**Figure 5.** One distillation step. Encode the view. Add noise. Run the frozen UNet one time. The residual becomes the gradient. It reaches the prime weights only.

**Figure 6.** Every symbol the figures use, with the value or shape it carries and the function that computes it.

Two phases

Phase 2 is Dream Target. It uses DreamShaper LCM. Each round asks img2img for a cleaner target. Then SSIM and MSE pull the derived views toward that target. Strength starts high and decays to a light polish.

The optimizer CLI defaults are 500 SDS steps and 8 Dream rounds. We baked the gallery with a research recipe. It uses 5000 SDS steps and 1 Dream round. That recipe is not only a step budget. It also swaps the prime network, the phase 1 optimizer, the guidance, and the SDS objective. Read every number on a figure with that in mind.

The loop creates a fresh Adam optimizer at the phase boundary. SDS gradients are about four orders of magnitude larger than Dream Target gradients. Without the reset, phase 2 moved loss by under 2 percent per round. With the reset, loss moved 44 to 60 percent. Those figures come from a short flip run. The run used 250 SDS steps and 4 Dream Target rounds of 150 steps each.

**Figure 7.** Real checkpoints from the calibration smoke run. SDS 250 shows noise. SDS 5000 shows a giraffe. The loss swings sign the whole time, so it cannot rank quality.

**Figure 8.** A real round-1 triple. It shows the derived view, the dreamed target, and the regressed view.

Joint Dream

Independent Dream Targets can fight over the same pixels. Joint Dream denoises both flip views together. It averages them in the upright frame. Then it steps from that consensus. The two views become two orientations of one image. Pixel space is required for this step because the SD 1.5 VAE does not commute with a 180-degree turn.

The loop averages in pixel space. The SD 1.5 VAE does not commute with a 180-degree turn in latent space. Joint Dream is an opt-in flag. It is not the product default. It can rescue a pair whose shapes can be one picture. It can also collapse a pair whose subjects cannot share one picture.

**Figure 9.** Both views denoise together and reconcile to one consensus image in pixel space. Independent targets disagree. Joint targets are one image seen two ways.

What we actually measured

A good result is a specific pair, seed, and viewing mode. It is not a recipe that works every time. We pick keepers by eye. An automatic CLIP screen failed. Its ROC-AUC was 0.706 against a 0.75 bar. So we do not use it.

Prompt wording is the biggest lever. Arguments do not predict it. Frames and extra objects come from the specific phrase. They do not come from sketch versus oil as a medium. We kept oil for color and for fewer photographic frames. We did not keep oil because it won on yield.

We rejected negative prompts. Extra Dream rounds after the first did not help: in one run, readable images fell from 35 to 25 of 182, though a later test did not clear the bar we set to call the drop real. 256 px primes were enough. 512 px cost about 3.3 times as much for no visible gain. That is the prime render resolution, not the Score Distillation ladder.

These conclusions came after the gallery was baked. The 14 keepers in the gallery span both wordings and both styles. 11 of 14 used joint Dream. The rest used independent targets. Export window2-2026-08-clean.

**Figure 10.** The shipped defaults against the recipe that baked this gallery. The recipe changes the network, the optimizer, the guidance, and the objective.

**Figure 11.** 206 cells, 43 at score four or five, 26 keepers. A frame defect cut 16 of them, and one was never frame-rated. CLIP scores are recorded. They are not a gate.

**Figure 12.** One pair at three seeds, baked twice. Joint Dream on top, independent targets below. Nothing else changes, and the human score runs from zero to five.

**Figure 13.** Three real cells that did not become keepers, each with both views and the reason it was cut.

How to print it

For flip, print the prime on paper. Turn the sheet around. Laser output is enough for a desk test. Rotate and hidden types need transparency film and a backlight. They are not on this page.

**Figure 14.** Print the prime. Then turn the sheet 180 degrees.

Not in the studio yet

There is no illusion job in the API. There is no designer in the app. The optimizer is a worker CLI on an unmerged branch. This page is the public study of that research.

When jobs land, they will sit behind realtime drawing sessions. Other batch work sits there too. Until then, the live product is still the canvas loop in the whitepaper.

A typeset paper can follow. This page is the public study until then.