PanoDiff-SR: Synthesizing Dental Panoramic Radiographs using Diffusion and Super-resolution

1Before anything else, take the test

Six dentists sat the same test on these exact images. Twelve seconds per radiograph, one hundred real and one hundred synthetic, no second look. Their mean accuracy was 68.5%. Chance is 50%.

The images below are drawn from that study, in the resolution the observers saw. Twelve are real radiographs from the pooled public corpus; twelve came out of PanoDiff-SR. After each answer you see the ground truth and, more interestingly, what each of the six clinicians said about that same image.

Real or synthetic?

Live Keyboard: R for real, S for synthetic, → for the next image.

1 / 24

Twelve seconds are up. Answer from memory — that is the decision the six dentists were making.

–your accuracy

0images answered

What the six dentists said about this image

Answer first — the responses appear here.

Benchmarks

You, so far–

Six dentists, this study68.5%

Readers vs. StyleGAN2-ADA PRs78.2%

Guessing50.0%

Your score is over 24 images, theirs over 200, and you are free to take as long as you like. It is a demonstration, not a measurement.

68.5%

Mean accuracy of six dentists over 200 images, with unsure responses weighted as half a correct and half an incorrect decision.

κ = 0.18–0.43

Pairwise agreement between observers, which is fair at best. They were not converging on the same tell-tale features.

157

Judgements the observers refused to commit on. They are the hardest images, and how you count them is the one thing that moves the headline number.

2Why it is built in two stages

A panoramic radiograph is a wide, thin image of the whole jaw. Generating one directly at diagnostic resolution with a diffusion model is possible and expensive. PanoDiff-SR does not.

Machine learning in dentistry is short of data. Public panoramic radiograph (PR) datasets are few, each is tied to one device and one patient population, and none of them contains many rare findings. Synthetic radiographs are a way out of that: pooled with real ones they can balance an under-represented class, and on their own they make teaching material that carries no patient inside it.

Until now that work has been almost entirely GAN-based, and three problems keep recurring. The field of view is usually cropped to the dentoalveolar region rather than the complete arch. Observer studies still catch the synthetic images at around 78–82%, which means visible artefacts survive. And adversarial training remains fragile — mode collapse and instability are reported repeatedly in medical image synthesis.

Diffusion answers the third problem directly. It does not answer the second one for free, and it makes the first one expensive: the iterative denoising loop has to run at whatever resolution you want out, and at 1024 × 512 that cost is prohibitive on one GPU. The pipeline therefore splits the two jobs. Generation happens small; resolution is added afterwards in a single pass.

The pipeline, end to end

Paper §3, Table 2 Click a stage.

What each stage costs

The split is only worth making if it is cheaper, so the paper measures it. Both stages were profiled back to back on the same single 47 GB GPU.

Computational cost of the two stages

Paper Table 2

	PanoDiff (LR synthesis)	SR — HAT 4× (LR → HR)

* wall-clock from the training logs of the runs reported in the paper. † from a dedicated profiling run measuring both stages back to back. LR: 256 × 128. HR: 1024 × 512.

250 : 1

Network evaluations per image. The generator runs its denoiser 250 times; the super-resolution network runs once.

7.15 s

End-to-end inference per radiograph, of which 6.84 s is the diffusion loop at 256 × 128 and 0.31 s is the 4× upscale.

1 GPU

Everything in this paper — both trainings and all 7243 generated radiographs — ran on one RTX 6000 Ada.

3Five datasets, five devices, one corpus

7243 panoramic radiographs pooled from five public sources on four continents, cropped per source and resized to a common 1024 × 512.

7,243radiographs after preparation

paper §4

5source datasets, five different devices

paper Table 1

5countries: China, Switzerland, DR Congo, USA, Brazil

paper Table 1

1024×512common resolution every image is resized to

process_data.py

The five sources

Paper Table 1 Click a row for what it contributes.

Source	Images	Country	Native resolution

Pooling five devices is not free. Each machine has its own exposure, geometry and post-processing, and a t-SNE of ResNet-50 features shows the five sources occupying distinguishable regions of feature space. The model therefore learns a mixture of device appearances, and every distribution metric in this paper is computed against that mixture.

t-SNE of 500 random images from each of the five source datasets, forming distinguishable clusters — **Five devices, five neighbourhoods.** t-SNE of 500 random images per source dataset. The clusters separate, which is exactly why the paper later stratifies its FID by device (§8) rather than reporting one pooled number and stopping.

What preparation does to an image

process_data.py

One step, applied to everything: a crop that is hard-coded per source to remove that device's letterboxing and burned-in markers, then a Lanczos resize to 1024 × 512. The crop margins are the whole trick — without them, five differently framed corpora cannot be pooled into one training set.

crop_top_left_bottom_right, per source
  ADLD     (90, 30, 30, 170)
  DENTEX   (20, 100, 50, 20)
  TUFTS    (20, 30, 100, 100)
  USPFORP  (30, 30, 165, 30)
then  img.crop(box).resize((1024, 512), LANCZOS)

4Stage one — PanoDiff

A 34-million-parameter U-Net, trained to strip noise out of a radiograph, then run backwards 250 times from pure noise until a jaw appears.

Diffusion has two halves. The forward half needs no learning at all: take a real radiograph and add Gaussian noise to it on a fixed schedule until, after 1000 steps, nothing is left but noise. The reverse half is the model: a U-Net trained to look at a noisy image and say how much of it is noise. Run that network repeatedly, subtracting a little of its prediction each time, and you walk backwards from noise into an image.

The forward process, on a real radiograph

Illustration Drag the timestep. The schedule is the cosine one the model was trained with.

t = 0t = 250t = 500t = 750t = 1000

Timestep t0

Signal retained

√̄ᾱ_t, the weight on the image1.000

√̄1−ᾱ_t, the weight on the noise0.000

x_t = √(ᾱ_t) · x_0 + √(1 − ᾱ_t) · ε, ε ~ N(0, I)

The two schedules are drawn to their real definitions, so the switch shows a genuine difference: the cosine schedule destroys the image later and more gently. That choice turns out to matter, and not in the direction the paper originally assumed — see the ablation below.

The three phases, as the paper draws them

Network

U-Net with hidden dims [64, 128, 256, 512] and self-attention, 34.0 M trainable parameters.

Downsampling

Only two spatial halvings, by pixel-unshuffle rather than strided convolution, so the coarsest feature map is 32 × 64 and the reduction moves information into channels instead of discarding it.

Training

L₁ objective, 1000 training timesteps, cosine β-schedule, EMA weights, 110 epochs, seed 42.

Sampling

DDIM with 250 inference steps — far fewer than ancestral DDPM sampling would need for the same quality.

Diagram of the reverse phase: a trainable U-Net predicting noise, contrasted with step-by-step DDPM denoising — **The reverse phase.** Top: the U-Net (flame — trainable) sees a pair (x₀, x_t=k) and is trained with an L₁ loss to predict the noise in x_t=k. Bottom, for contrast: step-by-step DDPM denoising with the same network frozen (snowflake), for which a single-step estimate is not adequate at high noise levels. Paper Figure 5.

Watch it learn

Because the sampling seed is fixed, the same noise vector can be decoded by each checkpoint in turn, so what changes between columns is the generator alone. The striking thing is how early the coarse structure arrives: by epoch 11 the model already draws a recognisable full arch. What the later epochs add is consistency and detail — firmer crown outlines, roots that separate from the trabecular background, and even contrast across the image.

The same five seeds, across training

Checkpoints ep11 → ep110 Drag the epoch, or press Play. Click any radiograph to enlarge.

Epoch 11

Epoch11

All five columns are upscaled by the same super-resolution model, so what changes between epochs is the generator alone.

What the diffusion stage is actually sensitive to

Four things were varied, three of them read straight off the trained checkpoint and one of them — the noise schedule — requiring two models trained from scratch under a matched 30-epoch budget. FID here is computed on 1000 generated against 1000 real low-resolution radiographs, so these numbers are comparable with each other but not with the full-corpus tables further down.

Diffusion ablation

Paper Table 10 Hover a point. The axes trade quality against sampling cost.

Weight averaging DDIM sampler budget Noise schedule, matched 30-epoch runs

Two findings the authors report against themselves.

Weight averaging is doing a great deal of work. Reading the same checkpoint from raw instead of EMA weights raises FID from 67.8 to 122.5. The Inception score moves the other way (2.65 → 2.78), which is a clean reminder that IS rewards confident, varied ImageNet predictions without reference to the real distribution at all.
The cosine schedule was the wrong default. At a matched 30-epoch budget a linear schedule reaches FID 49.0 against the cosine schedule's 80.9. The cosine schedule was inherited from the reference implementation and never tuned. Every result on this page comes from the cosine model; the paper reports the negative finding rather than quietly switching.

The sampler budget is the cheap lever: 100 steps costs 5 FID points and runs 2.5× faster, 50 steps stays within 15 points at 4.9× faster, and only at 25 steps does quality fall away.

Diffusion against a GAN, at the resolution both were trained at

The comparison that matters for stage one is at 256 × 128, before super-resolution confuses things. FastGAN was trained on the same corpus.

Low-resolution samples, side by side

Paper Figure 14 Click any sample to enlarge.

PanoDiff — FID 55.0, IS 2.99

FastGAN — FID 85.2, IS 2.25

FID is against the same 7243 real low-resolution radiographs; real data itself scores IS 2.91. The confound is stated in the paper and worth repeating here: the two models do not match in capacity, training budget or pre-training, so this figure shows what each configuration produced, not which family is better. The controlled version of that question is in §8.

5Stage two — super-resolution

A hybrid attention transformer, pre-trained on natural images and fine-tuned for 400,000 iterations on radiograph pairs, takes the 256 × 128 seed to 1024 × 512 in a single forward pass.

The seed the diffusion model produces is a complete panoramic radiograph in composition but useless in detail: at 256 × 128 a molar is a smudge a dozen pixels across. HAT (hybrid attention transformer) combines channel attention with overlapping window self-attention, so it can use both local texture and long-range context; here it starts from released natural-image weights and is fine-tuned on PR pairs with pixel, perceptual and adversarial terms.

The seed and the result

Generated pairs Drag across the image. Left is the 256×128 seed, bicubic-upscaled for display.

The same radiograph after HAT super-resolution

The 256 by 128 diffusion seed, bicubic-upscaled

LR seed, bicubicHAT-SR, 1024×512

↔

Both halves are the same generated radiograph. The left is what stage one emitted; the right is what stage two made of it.

What the upscale adds

Enamel edges, the periodontal ligament space, trabecular texture in the ramus and the outlines of restorations. What it cannot add is anything the seed did not imply — the anatomy is decided at 256 × 128.

Does fine-tuning earn its 53 hours?

Measured honestly, on 300 real radiographs downsampled 4× and restored, the answer depends entirely on which metric you believe — and this is the clearest case on the page of a metric disagreeing with the eye.

Super-resolution ablation

Paper Table 9 Switch the metric. Bars are scaled within each metric.

Bicubic upsampling wins on PSNR and SSIM and is visibly the worst image on the page. PSNR and SSIM reward a smooth, conservative prediction, and blur is exactly that. LPIPS, which compares deep features rather than pixels, ranks the same rows in the opposite order — and it is LPIPS that agrees with what the crops below look like. Fine-tuning takes LPIPS from 0.274 to 0.116; the last of that comes from weight averaging.

Two magnified crops of one radiograph restored five ways: bicubic, stock HAT, fine-tuned without EMA, fine-tuned with EMA, and ground truth — **The same pixels, five ways.** Left to right: bicubic upsampling, the released HAT model with no fine-tuning, our fine-tuning without EMA, our fine-tuning with EMA, and the ground truth. All crops are the same region at the same magnification. Paper Figure 16.

HAT against SwinIR

The first version of this comparison was not fair, and the revision says so. SwinIR and HAT were compared without equalising pre-training or fine-tuning budget. Under matched conditions the picture changes: before any fine-tuning the two architectures are perceptually equivalent (LPIPS 0.274 against 0.269), and what separates them afterwards is the objective, not the architecture — HAT was fine-tuned with perceptual and adversarial terms, the recovered SwinIR with a pixel loss alone.

Matched pre-training, matched budget

Paper Table 15

Variant	PSNR ↑	SSIM ↑	LPIPS ↓

SwinIR fine-tuned on a pixel loss posts the best PSNR and SSIM on the page and the worst LPIPS — the same trap as bicubic, one row further along. Fine-tuning SwinIR with HAT's exact recipe is the experiment that would settle the architecture question, and it has not been run.

Four upscaling arms on the same seeds

Paper Figure 15 Pick a row, then switch the arm.

A synthetic radiograph produced by the selected generation and upscaling arm

6How close is the distribution?

Four sets of 7243 images — real and synthetic, at low and high resolution — and every pairwise comparison the paper could make between them.

The headline is FID 40.7 between 7243 real and 7243 synthetic high-resolution radiographs. On its own that number means very little, which is why the table below includes the comparisons that give it a scale: what two halves of the same set score against each other (the floor), and what two different real devices score against each other (the ceiling of what "different but both real" looks like).

Every FID comparison in the paper

Paper Table 3 Hover a bar. Lower is more similar.

Self-comparison — the floor Real vs. synthetic Real device vs. real device High vs. low resolution

Read the rose bar at 40.7 against the green band at 36.8–70.1. The distance from real to synthetic at high resolution sits inside the range that separates two real datasets acquired on different machines. Read it also against the grey floor at 5–6, which is what the same distribution scores against itself: 40.7 is not small in absolute terms.

t-SNE of real and PanoDiff features at low resolution, left, and high resolution, right — **Super-resolution pulls the two clouds together.** t-SNE of ResNet-50 features for real (GT) and PanoDiff (PD) radiographs. At low resolution, left, the clusters are adjacent but distinct and their centroids are 81.41 apart. At high resolution, right, the synthetic cluster sits surrounded by real points and the centroid distance falls to 34.73. Paper Figure 8.

Inception scores

Paper Tables 4 and 7 Switch the view.

What these two metrics can and cannot say. FID compares the mean and covariance of Inception features between two sets; it is sensitive to sample size, so numbers are only comparable within a table. Inception score never looks at the real data at all — it rewards an ImageNet classifier being confident and varied, which is why a model can improve on IS while drifting away from the distribution it is supposed to be modelling. That happens on this page, in the diffusion ablation. Neither metric was designed for radiographs, and neither knows what a mandible is. That is what the next two sections are for.

7Six dentists, 200 radiographs, twelve seconds each

Three early-career dentists with under ten years of practice and three with over fifteen, spanning prosthodontics, oral surgery, orthodontics and oral radiology. Each sat one session of about an hour, alone.

Each radiograph was displayed for twelve seconds, including about two seconds of loading. For each one the observer moved a slider to one of five positions: definitely real, probably real, unsure, probably fake, definitely fake. The graded scale is deliberate — forcing a binary choice would have thrown away the information that matters most, which is how often an expert simply could not tell.

How each observer used the scale

Raw responses, 6 × 200 Click an observer. Bars are the raw counts, no weighting applied.

Derived metrics, certainty-weighted

Bars read left to right: definitely real, probably real, unsure, probably fake, definitely fake. The counts here are computed from the study's raw submission log and reproduce the paper's Table 5 exactly.

The one convention that moves the number

A five-level scale needs a rule for what an unsure response is worth, and it is fair to ask whether that rule is carrying the headline result. The paper reports all four candidate rules.

Mean accuracy under four conventions

Paper §4.1 Click a convention.

Three of the four agree to within 3.4 points, so the reported accuracy is a property of the responses rather than of the weighting. The fourth — discarding unsure responses — lifts the mean to 78.5% and lands it exactly on the 78.2% GAN benchmark the paper's central claim rests on being below. It does that not because the radiographs became easier to detect but because the 157 hardest judgements were removed from the denominator.

ROC and precision–recall, per observer

Computed from the raw responses Hover a curve. Toggle a group in the legend.

They did not agree with each other either

Weighted Cohen's κ between every pair

Paper Table 6 Quadratic weights, so a near-miss is penalised less than a reversal.

Values run from 0.18 to 0.43, which is fair at best on the Landis–Koch scale, and there is no meaningful difference between the early-career and experienced groups. In a real-versus-synthetic task this is the favourable direction: high agreement would mean the observers were all spotting the same artefact. Low agreement, together with 68.5% accuracy, says a sizeable fraction of the synthetic radiographs carried no consistent tell at all. The highest pair, EC3 and EP3 at 0.43, is also the pair whose response distributions look most alike.

What each kind of mistake looked like

Paper Figure 12 Pick a cell of the confusion matrix, then a certainty level.

Examples were chosen for each category by the majority of the observers' assessments, not by eye. Click either image to enlarge.

Six pie charts, one per observer, splitting decisions into correct and incorrect at full and partial certainty — **Every decision, per observer.** Correct and incorrect decisions split by whether the observer was fully or partially certain. The size of the incorrect-but-fully-certain wedge is the interesting part. Paper Figure 9.

8What a machine sees that a dentist does not

A vision transformer trained to separate real from synthetic reaches 97.7% test accuracy. The humans reached 68.5%. Where the classifier looks is a more delicate question than the paper first claimed.

A ViT was trained on 7243 real and 7243 synthetic high-resolution radiographs with an 80:20 split. It separates them almost perfectly, which is worth stating plainly: these images are not indistinguishable, they are hard for a person under time pressure. Attention rollout then shows which patches drove each decision.

Attention map viewer

Paper Figure 13 Pick a sample, then slide between the radiograph and its map.

Radiograph ↔ map55%

Top three rows of the grid are real radiographs, bottom three synthetic. A dot marks the four the classifier got wrong; all four are synthetic radiographs it read as real. C is the classifier's output, 1 for real and 0 for synthetic.

A claim withdrawn in revision. The submitted version of this paper said attention on synthetic images was misdirected and did not concentrate on the same anatomical cues. Measured over the full held-out set, that does not hold — synthetic maps put marginally more mass on teeth and jaw than real ones do. The sentence was withdrawn and replaced by the measurement below, which also found something the authors were not looking for: this classifier does not preferentially attend to dental anatomy in either class.

Attention concentration over the full held-out set

Paper Table 13 1449 real and 1449 synthetic radiographs. Hover a row.

The dashed marker on each row is the uniform reference: the value the quantity would take if attention carried no spatial preference at all. Both classes fall below it on teeth and on jaw, which means the network is attending away from the dentition, not towards it. With 1449 images per group every difference here is significant; the effect sizes are the quantity to read, and only entropy (+0.607) is large.

Mean attention maps over all held-out real and synthetic radiographs, and their difference — **Averaged over every held-out radiograph.** Solid contour: the maxillomandibular prior. Dashed contour: the teeth prior. Colour is in multiples of the uniform value and the two mean maps share a scale, so their relative peakedness is comparable. Structure survives the averaging — the real mean map peaks at 2.55× uniform against 1.80× for the synthetic — but the peak sits on the inferior image border, outside the jaw, rather than on the teeth. Paper Figure 18.

9Is the two-stage split actually the right call?

Four models trained from scratch at full resolution on the same 7243 radiographs under a matched budget, to test the premise the whole pipeline rests on. The answer is not a clean win.

If splitting generation from upscaling is the contribution, then generating directly at 1024 × 512 is the control that has to be run. It was: StyleGAN2-ADA, FastGAN-HR, ADM, LDM and PanoDiff-HR — the same model as PanoDiff but trained at full resolution — all on the same data, all at 110 epochs.

Direct high-resolution synthesis against the two-stage pipeline

Paper Table 11 Hover a point. Down and left is better on both axes.

Diffusion GAN Two-stage (ours) Ring size ∝ training cost

Model	Family	FID ↓	IS ↑	ViT detect % ↓	Cost (acc.-h)

Three things this table settles, one of them against the paper.

The split works. PanoDiff-HR — the identical model trained directly at full resolution — scores FID 155.4 against the two-stage pipeline's 40.5, on more training compute. Generating small and upscaling is not a compromise here, it is the thing that makes the model work at all.
StyleGAN2-ADA beats it on fidelity. FID 23.5 against 40.5, and it is the only arm the real-versus-synthetic classifier finds genuinely hard (89.6% against 97.6%). It cost 441 accelerator-hours to get there. The paper's claims are therefore stated in terms of cost and of observer detectability, not of image fidelity.
No model memorised its training set. A nearest-neighbour test in Inception feature space found no generated image closer to a training radiograph than real radiographs typically lie to each other. That bounds whole-image copying; it would not catch the reuse of one distinctive local structure.

Does it look like a jaw?

FID and IS have no idea what a mandible is, so four global anatomical measures were computed over 1500 images per model: bilateral symmetry, the periodicity of the tooth row, the number of distinct enamel crowns, and occlusal separation.

Anatomical plausibility

Paper Table 12 The real column is the target, not a winner.

* differs from real at p < 0.05, ** at p < 10⁻⁴ (Mann–Whitney U). The last row is the fraction of images falling outside the central 98% of real radiographs on any of the four measures, so the real column's 7.3% is the floor that statistic can reach. LDM's 53.9% and its 26.1 enamel crowns are the signature of the repeated tooth rows visible in the figure below. These measures reject degenerate output reliably; they are global, and they say nothing about crown-to-root proportion or root morphology.

Random samples from each direct high-resolution baseline alongside real radiographs and the two-stage pipeline — **Random samples, fixed seed, not chosen by eye.** PanoDiff-HR, which differs from PanoDiff-SR only in generating at full resolution, is blurred and sparse. The latent model repeats tooth rows. StyleGAN2-ADA is sharp and, as the table above records, is not separated from real radiographs by these global measures at all. Paper Figure 17.

Is the model closer to one device than another?

The corpus pools five machines, so a fair worry is that the model has simply learned the largest source. Each device was therefore compared with the synthetic set and with the pooled real corpus at a matched n = 500, redrawn five times.

Device-stratified FID

Paper Table 14 Hover a pair of bars.

to the synthetic set to the pooled real corpus

The two bars move together: a device's distance to the synthetic set is predicted by its own distance to the pooled corpus, so the model is not closer to one machine than another relative to how far the machines sit from each other. For scale, the ten device-to-device comparisons range from 66.6 to 123.4, and two disjoint draws from the pooled real set give 56.3 ± 1.7 at n = 250. What this does not resolve is how much of the absolute level, as opposed to the ordering, is device appearance rather than anatomy.

10What this does not show

The limitations the authors put in writing, rather than the ones a reader has to infer.

Sampling is unconditional

Neither anatomy nor pathology can be specified. That is a real limit on the use case that motivates synthetic data in the first place — augmenting rare findings — and it is the obvious next piece of work.

The configuration is documented, not optimised

U-Net width and depth, the SR loss weights, the number of training timesteps and the composition of the pooled corpus were adopted from the source implementations and never varied.

The schedule comparison was short

Linear versus cosine was run at 30 epochs, not the 110 of the reported model. A longer budget could narrow or widen the gap.

Six readers, one session each

No repeated readings, so within-reader variability is unknown. With three readers per group the early-career versus experienced test cannot reach p < 0.10 by construction, and the observed gap does not reach significance.

The SR architecture question is open

Matched pre-training and budget removed two confounds, but HAT was fine-tuned with perceptual and adversarial terms and the recovered SwinIR with a pixel loss. The objective, not the architecture, is what produced the difference.

Memorisation is bounded, not excluded

The nearest-neighbour test rules out whole-image copying. It would not detect a distinctive local structure being reproduced, which matters because the source data, though public, comes from patients.

The claim is not that these radiographs are indistinguishable from real ones. A classifier separates them at 97.7%. The claim is that six clinicians, reading at the speed clinicians read, reached 68.5% — and that this was done on one GPU.

11Code, models and citation

Both stages, the observer-study application the dentists actually used, and the checkpoints behind every figure on this page.

Repository map

github.com/s4nyam/PanoDiff Pick a directory.

Diffusion checkpoints

Six epochs of PanoDiff — 11, 33, 55, 77, 99 and 110 — the six columns of the training progression above.
archive.org/download/panodiff/trained_models

Super-resolution weights

Real_HAT_GAN_SRx4_finetuned, the 400k-iteration EMA model used for every high-resolution image on this page.
archive.org/download/panodiff/experiments

Source datasets

All five are public or available on request. Links and terms are in §3; none of them is redistributed here.

Cite this work

@misc{jain2025panodiffsr,
  title         = {PanoDiff-SR: Synthesizing Dental Panoramic Radiographs using Diffusion and Super-resolution},
  author        = {Jain, Sanyam and Neves de Freitas, Bruna and Basse-O'Connor, Andreas
                   and Iosifidis, Alexandros and Pauwels, Ruben},
  year          = {2025},
  eprint        = {2507.09227},
  archivePrefix = {arXiv},
  primaryClass  = {eess.IV},
  url           = {https://arxiv.org/abs/2507.09227}
}

Copied Preprint: arXiv:2507.09227