ImageDDI: Deep Interactive Image Workflows

• ImageDDI is an image-centric, human-in-the-loop framework that generates adjustable image states for tasks like CT denoising and diffusion inversion.
• It enables real-time test-time control by letting users navigate noise–resolution tradeoffs in low-dose CT and achieve exact, reversible diffusion trajectories.
• The approach integrates controllable denoising with reversible editing mechanisms, enhancing flexibility and reconstruction accuracy in both clinical and creative imaging applications.

ImageDDI denotes an image-centric class of deep, interactive workflows in which a model exposes an adjustable family of image states at test time rather than a single fixed output. In the cited literature, the term does not appear as a single standardized algorithmic name. In low-dose CT, it is explicitly aligned with Deep Interactive Denoiser (DID), an image-centric, CT denoising framework in which “Image” denotes image-centric CT denoising, “D” denotes deep learning, “D” denotes denoising / diagnosis driven, and “I” denotes interactive, user-controllable operation (Bai et al., 2020). In diffusion-model research, the same label is discussed contextually in relation to image-based DDIM inversion and editing pipelines, where real images are mapped into deterministic diffusion trajectories and then edited under prompt-conditioned guidance, as in BDIA and SAGE (Zhang et al., 2023, Gomez-Trenado et al., 14 May 2025). This suggests that ImageDDI is best understood as a family resemblance across image-centric, human-in-the-loop systems built around controllable denoising or controllable diffusion inversion.

1. Terminological scope and research domains

The term spans two technically distinct research settings. In CT, the emphasis is on low-dose image restoration with an explicit noise–resolution tradeoff. In diffusion editing, the emphasis is on inversion consistency, round-trip reconstruction, and prompt-based modification of real images. The shared theme is not a common backbone or loss, but a common interaction model: the system produces multiple plausible image states and permits test-time control over which state is selected.

Usage of ImageDDI	Core task	Representative work
DID-style ImageDDI	Interactive LDCT denoising with noise–resolution control	DID (Bai et al., 2020)
Diffusion-style ImageDDI	Image-based DDIM inversion and editing	BDIA (Zhang et al., 2023), SAGE (Gomez-Trenado et al., 14 May 2025)

A frequent misconception is to treat ImageDDI as a single canonical method. The cited sources do not support that interpretation. Instead, one source explicitly maps the term onto DID as a deep, interactive denoising interface for CT (Bai et al., 2020), while two diffusion papers discuss it only contextually as image-based DDIM inversion or image editing workflows (Zhang et al., 2023, Gomez-Trenado et al., 14 May 2025). The result is a polysemous term whose meaning is determined by domain: NRT navigation in medical imaging, or inversion-guided controllable editing in latent diffusion.

2. ImageDDI in low-dose CT: the Deep Interactive Denoiser

In LDCT, reducing the X-ray dose increases quantum noise, which can obscure low-contrast lesions and subtle pathology, compromise guidance for interventions or radiotherapy, and reduce radiologists’ confidence. DID was introduced to address two limitations of standard deep denoisers: a trained model normally produces only one denoised output per input, and its generalizability degrades when testing noise levels differ from those present in training (Bai et al., 2020).

The baseline supervised denoising model is a CNN $\phi_\theta$ trained from noisy LDCT images $y_m$ to clean NDCT counterparts $x_m$ by minimizing

$$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$

At test time, standard inference yields a single image,

$\bar x = \phi_{\theta_0}(y),$

which fixes the learned noise–resolution balance for every case (Bai et al., 2020).

DID converts this fixed mapping into a test-time adjustable process. For an LDCT input $y$, it defines two bound images. The high-resolution / high-noise bound is the original input,

$x_{\mathrm{high}} = y,$

and the low-noise / low-resolution bound is obtained by recursively applying the pre-trained denoiser $K$ times,

$x_{\mathrm{low}} = \phi_{\theta_0}^K(y).$

In the reported experiments, $K=3$ was chosen empirically as the smallest value yielding an “over-denoised” image that remains clinically meaningful (Bai et al., 2020).

Starting from $y_m$0, DID performs single-sample fine-tuning toward each bound:

$y_m$1

$y_m$2

This produces sequences $y_m$3 that move monotonically toward smoother, lower-noise outputs or toward sharper, higher-noise outputs, thereby spanning an NRT curve (Bai et al., 2020). In the NTK regime, the update is written as

$y_m$4

with $y_m$5 or $y_m$6, so each step partially corrects the residual and drives the image exponentially toward the chosen bound.

The main architecture used in the study is a 2D U-Net with input size $y_m$7, downsampling depth 9, initial feature channels 32 doubled at each downsampling step, $y_m$8 convolutions with instance normalization and ReLU, stride-2 convolutions for downsampling, and the standard U-Net mirror with feature concatenation for upsampling. Training uses Adam with initial learning rate $y_m$9, reduced to $x_m$0 and $x_m$1 at $x_m$2 and $x_m$3 iterations, $x_m$4, $x_m$5, batch size 1, and total $x_m$6 iterations (Bai et al., 2020).

3. Noise–resolution navigation, datasets, and generalizability

The interaction model of DID is intentionally simple. The paper describes a slider or knob with one extreme labeled “High Resolution” and the other labeled “Low Noise”; internally, this corresponds to selecting an iteration index $x_m$7 in the optimization sequence. Because the U-Net processes $x_m$8 images and each iteration takes on the order of tens of milliseconds on a single NVIDIA Titan X GPU, the system is described as real-time in the denoising context, either through on-the-fly generation or pre-computed candidate sets (Bai et al., 2020).

The noise–resolution tradeoff is evaluated quantitatively using standard deviation in a flat region, RMSE against NDCT, and lesion CNR,

$x_m$9

where $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$0 are lesion and background means and $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$1 are their standard deviations. As reported, STD decreases toward the low-noise side, while visual detail increases toward the high-resolution side (Bai et al., 2020).

The validation program emphasizes cross-noise and cross-domain robustness. On the AAPM LDCT Challenge dataset, DID yields multiple NRT candidates from a single input; low-noise images improve CNR for low-contrast lesions, while high-resolution candidates reveal high-contrast point-like structures more clearly. On a cadaver study acquired on a GE scanner at noise indexes $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$2, a denoiser trained only on simulated 25%-dose Siemens data over-smoothed the relatively clean NI=10 and NI=20 cases, while DID’s high-resolution direction restored structures such as curved anatomical features. On a CBCT dataset from a Varian TrueBeam radiotherapy linac, again using training from the AAPM dataset, higher-dose CBCT images were over-smoothed by the default denoiser, whereas DID recovered resolution or suppressed noise as needed (Bai et al., 2020).

The same paper also reports a controlled mismatch experiment across synthetic dose levels $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$3, $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$4, $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$5, and $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$6 relative to NDCT. Networks trained with higher-dose data tended to under-denoise a $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$7-dose test image, and DID’s low-noise direction corrected that. Networks trained with lower-dose data over-denoised the same test image, and DID’s high-resolution direction recovered details. Even when the training dose matched the test dose, DID still added flexibility around the default optimum (Bai et al., 2020).

Architecture dependence was also examined. A plain network with 8 stacked convolution–instance-norm–ReLU layers and no U-Net structure still produced a monotonic NRT under DID, confirming that the framework is not tied to a particular denoiser topology (Bai et al., 2020). This is central to the CT interpretation of ImageDDI: it is a test-time interface layered on top of an existing deep denoiser, not a replacement for the denoiser itself.

4. ImageDDI in diffusion models: DDIM inversion and exact reversibility

In diffusion-model image editing, ImageDDI refers contextually to image-based DDIM inversion workflows. The core problem is that a real image must be mapped into a latent trajectory so that it can be edited and then denoised back to image space with minimal drift. Standard DDIM inversion is only approximate, because the inverse step replaces the unavailable $$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$8 with an estimate evaluated at a different state, which causes round-trip inconsistency and loss of detail under editing (Zhang et al., 2023).

BDIA introduces a bi-directional integration approximation for exact diffusion inversion with negligible computational overhead. The deterministic DDIM update is

$$\theta_0 = \arg\min_\theta \sum_{m=0}^{M-1} \left\| \phi_\theta(y_m) - x_m \right\|_2^2.$$9

with

$\bar x = \phi_{\theta_0}(y),$0

BDIA refines this by combining a forward DDIM increment over $\bar x = \phi_{\theta_0}(y),$1 with a backward DDIM increment over $\bar x = \phi_{\theta_0}(y),$2, both computed from a single network evaluation at $\bar x = \phi_{\theta_0}(y),$3 (Zhang et al., 2023).

For $\bar x = \phi_{\theta_0}(y),$4, the central BDIA-DDIM update is

$\bar x = \phi_{\theta_0}(y),$5

Its significance lies in linearity: $\bar x = \phi_{\theta_0}(y),$6 is an affine function of $\bar x = \phi_{\theta_0}(y),$7 once $\bar x = \phi_{\theta_0}(y),$8 and $\bar x = \phi_{\theta_0}(y),$9 are fixed, so the reverse mapping is algebraic. The inverse is written explicitly in the paper and yields exact discrete-time recovery of $y$0 from $y$1, up to floating-point arithmetic (Zhang et al., 2023).

This exactness is meant in a restricted but important sense: the network is deterministic and fixed, the schedule is fixed, the same formulas and noise estimates are reused in forward and inverse directions, and no stochastic noise is added. Under those assumptions, BDIA-DDIM is a time-reversible map. Unlike EDICT, it does not require a second latent state and uses only one neural evaluation per step, giving negligible overhead over DDIM and roughly half the neural function evaluations of EDICT (Zhang et al., 2023).

The paper further reports that BDIA improves not only inversion but also sampling quality. On Stable Diffusion V2 with COCO2014 validation prompts, image size $y$2, and 10 timesteps, FID is 12.62 for BDIA-DDIM, 15.04 for DDIM, and 16.06 for the DPM-Solver configuration used there. On PNDM DDIM checkpoints, BDIA-DDIM improves FID over DDIM on CIFAR10 and CelebA across 10, 20, and 40 timesteps. The method is also extended to EDM through BDIA-EDM, where it produces consistent FID improvements on CIFAR10, FFHQ, AFHQV2, and ImageNet64 without extra neural evaluations (Zhang et al., 2023).

5. SAGE and inverse-DDIM-guided image editing

SAGE addresses a different limitation of diffusion-style ImageDDI: many real-image editing methods are either computationally intensive because they optimize per-image inversion variables, or they reconstruct poorly because inversion is too crude. SAGE uses inverse DDIM not primarily for exact trajectory recovery, but to collect semantically informative self-attention maps that can guide editing without explicit full-image reconstruction (Gomez-Trenado et al., 14 May 2025).

Given an image $y$3 and source prompt $y$4, the image is first encoded into a latent $y$5 and then inverted through deterministic DDIM to obtain the trajectory $y$6. During this inversion, SAGE records self-attention maps $y$7 from selected self-attention blocks of the diffusion U-Net. Editing then begins from the estimated noise latent $y$8 using a target prompt $y$9 (Gomez-Trenado et al., 14 May 2025).

The denoising direction is steered by a CFG-style mixture:

$x_{\mathrm{high}} = y,$0

with $x_{\mathrm{high}} = y,$1 in the reported experiments. Reconstruction of unedited regions is enforced through a self-attention loss,

$x_{\mathrm{high}} = y,$2

whose gradient with respect to $x_{\mathrm{high}} = y,$3 is used to guide the latent update. The paper describes a global self-attention guidance scale of 200 together with a decay schedule, following earlier time-dependent guidance practice (Gomez-Trenado et al., 14 May 2025).

SAGE complements this global structural guidance with cross-attention-based spatial control. Local blending uses an automatically inferred binary mask $x_{\mathrm{high}} = y,$4:

$x_{\mathrm{high}} = y,$5

so unedited regions are copied directly from the inversion latent. For word-swap edits, early-step cross-attention replacement can preserve the spatial footprint of the original subject, while cross-attention reweighting changes the strength of selected tokens (Gomez-Trenado et al., 14 May 2025).

Implementation details are explicit. The main experiments use Stable Diffusion 1.4 at $x_{\mathrm{high}} = y,$6 with 50 DDIM steps for both inversion and sampling; teaser examples use Stable Diffusion 2.1 at $x_{\mathrm{high}} = y,$7. For $x_{\mathrm{high}} = y,$8, self-attention maps of size $x_{\mathrm{high}} = y,$9 are taken from the second encoder block and the corresponding upsampling block, while cross-attention maps of size $K$0 from the second and third blocks are used for local blending. Local blending is active in the first 40 steps, cross-attention replacement in the first 5 steps, and a reweighting factor of 2.0 is used for certain tokens. On a single NVIDIA A100-40GB, SAGE FP16 adds about 12.6 seconds per $K$1 image and reaches 7.4 GB peak memory; the same supplementary comparison reports roughly 115 seconds and 21 GB for Null-text inversion (Gomez-Trenado et al., 14 May 2025).

Quantitatively, SAGE is presented as a balanced editing method rather than a pure reconstruction method. On PieBench, it achieves structure distance 11.0, background LPIPS 39.6, background SSIM 86.0, whole-image CLIP-T 25.5, and edited-region CLIP-T 22.0, giving the best background preservation and best whole-image CLIP-T among the listed baselines. On MagicBrush, it reports L1 6.4, L2 1.8, CLIP-I 90.9, DINO 85.9, and CLIP-T 27.6, with the best preservation metrics and near-best prompt fidelity. The user study covers 47 surveyed users in total, and the paper states that all differences are statistically significant at $K$2 using binomial tests and Fisher’s method for combining $K$3-values (Gomez-Trenado et al., 14 May 2025).

6. Conceptual synthesis, limitations, and open issues

Across both CT and diffusion contexts, ImageDDI is organized around test-time control. In DID, that control is expressed as navigation along a clinically meaningful NRT continuum defined by $K$4 and $K$5 (Bai et al., 2020). In BDIA, it appears as exact reversible traversal of a deterministic diffusion trajectory with a tunable parameter $K$6 (Zhang et al., 2023). In SAGE, it appears as prompt-conditioned editing constrained by inverse-DDIM-derived self-attention trajectories and spatial masks (Gomez-Trenado et al., 14 May 2025). This suggests a common abstraction: ImageDDI systems transform a static image model into an operator over controllable image manifolds.

The limitations, however, remain domain-specific. DID depends on the quality of the base denoiser, may in principle introduce subtle artifacts under over-optimization, does not provide a direct quantitative NRT indicator in the user interface, and was demonstrated primarily on 2D slice-based setups rather than full 3D volumes (Bai et al., 2020). BDIA is formulated for deterministic ODE solvers rather than stochastic samplers, depends on the accuracy of the score or noise predictor, is sensitive to the choice of $K$7, and claims exact inversion for BDIA-DDIM but not for BDIA-EDM (Zhang et al., 2023). SAGE is sensitive to the CFG scale and self-attention guidance schedule, can preserve unwanted structure during object removal because the reconstruction term conflicts with deletion, reports failure cases with over-saturated colors or poor reconstruction in about 6 of 700 PieBench examples, and performs worse in pure reconstruction settings where $K$8 than methods explicitly optimized for reconstruction (Gomez-Trenado et al., 14 May 2025).

A second misconception is that interactivity necessarily implies weaker quantitative performance. The cited work does not support that generalization. DID reports improved lesion CNR or better visibility of high-contrast structures depending on the chosen point on the NRT curve (Bai et al., 2020). BDIA reports better FID than DDIM in the tested regimes while also enabling exact inversion (Zhang et al., 2023). SAGE reports top or near-top results across PieBench and MagicBrush while remaining substantially lighter than optimization-heavy inversion methods (Gomez-Trenado et al., 14 May 2025).

The broader significance of ImageDDI is therefore methodological rather than terminological. In CT, it reintroduces to deep denoisers the tunability long associated with regularized iterative methods. In diffusion editing, it reinterprets inversion as a source of reversible states or semantic anchors, rather than merely a way to initialize noise. The unifying principle is a shift away from one-size-fits-all image generation and toward user-steerable, task-aware image computation.