RealDeal: Domain-Dependent Applications

• RealDeal is a domain-dependent term that, in biomedical imaging, denotes a 3D patch-based image-to-image diffusion refiner for brain MRI that restores sharp edges and fine anatomical details.
• In negotiation research, RealDeal-style systems refer to realistic dyadic bargaining agents that leverage modular strategies and retrieval-based natural language generation.
• The term is disambiguated from similar frameworks like DEAL, with studies reporting significant improvements in metrics such as LPIPS, FID, and KL divergence for refined image outputs.

Searching arXiv for the cited papers and topic variants to ground the article. RealDeal is a domain-dependent term with distinct meanings in current technical literature. As a named model, it denotes a 3D patch-based image-to-image diffusion refiner for brain MRI that operates in pixel space and is designed to restore sharp edges, fine textures, subtle anatomical features, and realistic imaging noise in latent-diffusion outputs (Zhu et al., 24 Jul 2025). In negotiation research, by contrast, “RealDeal” appears as a shorthand for realistic dyadic bargaining settings rather than as a standardized dataset name: FortUne Dial explicitly states that it does not use or reference a dataset or task named RealDeal, while modular negotiation work uses “RealDeal-style systems” to describe controllable, human-like bargaining agents (Sicilia et al., 2024, He et al., 2018).

1. Terminological scope and disambiguation

The term has at least three non-equivalent uses in the cited literature. The most explicit proper-name usage is the brain-image generation paper titled “RealDeal: Enhancing Realism and Details in Brain Image Generation via Image-to-Image Diffusion Models.” In negotiation work, the term is descriptive rather than canonical: one paper formalizes uncertainty-aware dialogue forecasting and then explains how its methods and metrics can be applied to a “RealDeal”-style dyadic bargaining benchmark, while another describes “RealDeal-style systems” as negotiation agents with controllable strategy and human-like language. A separate high-dimensional inference paper proposes DEAL, the “Debiased External-model-Assisted Lasso,” which is terminologically adjacent but distinct from RealDeal (Zhu et al., 24 Jul 2025, Sicilia et al., 2024, He et al., 2018, Zhang et al., 14 Jun 2026).

Usage	Domain	Status in the cited literature
RealDeal	Brain MRI generation	Named image-to-image diffusion refiner
“RealDeal”-style / “RealDeal-style systems”	Negotiation dialogue	Descriptive shorthand, not a standardized benchmark name
DEAL	High-dimensional regression	Distinct acronym, not RealDeal

A common misconception is to treat these as a single research object. The available sources instead support a disambiguated reading: RealDeal is a concrete generative model in biomedical imaging; in negotiation, it is a convenient label for realistic bargaining tasks and systems; and DEAL is a separate inferential framework.

2. RealDeal as post hoc diffusion refinement for brain MRI

In the biomedical setting, RealDeal addresses a specific limitation of latent diffusion models for brain MRI. The underlying claim is that latent diffusion models compress 3D volumes into a low-dimensional latent space using an autoencoder, and that this latent compression removes high-frequency content—sharp edges, fine textures, subtle anatomy, and realistic acquisition noise—causing oversmoothing and loss of anatomical fidelity. RealDeal formulates realism enhancement and detail addition as an image-to-image diffusion process applied after the latent diffusion model has produced a coarse 3D image (Zhu et al., 24 Jul 2025).

The method operates directly in pixel space rather than latent space. The latent diffusion model provides an initial coarse image, and RealDeal refines this image on 3D patches. Conditioning is twofold. First, the denoiser receives the LDM-generated coarse image patch, denoted $\hat{x}_{\text{patch}}$, which carries global anatomy and spatial context. Second, it receives a context patch $y_{\text{prev}}$ assembled from previously refined neighboring patches, intended to enforce spatial coherence and avoid grid artifacts. For the first, center patch, $y_{\text{prev}}$ is pure Gaussian noise. The patch schedule starts at the center patch and iteratively traverses neighboring patches until the entire volume is covered.

This design is explicitly targeted at restoring the realism missing from coarse latent reconstructions. The paper reports qualitative evidence that LDM reconstructions and LDM-synthesized images are smoother than real T1-weighted MRIs, particularly in edges and textures, whereas refined images recover sharp cortico-subcortical boundaries and realistic noise. A plausible implication is that the refiner is not intended to replace the coarse generator, but to act as a dedicated high-frequency reconstruction stage.

3. Mathematical formulation and implementation details

The forward process is standard DDPM diffusion in pixel space:

$$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$

with

$$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$

and

$$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$

The reverse model is conditional:

$$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$

parameterized by a 3D U-Net that predicts the noise residual $\epsilon$. The denoiser $f_\psi$ takes as input $[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$ and $y_{\text{prev}}$0, using explicit channel-wise concatenation for conditioning (Zhu et al., 24 Jul 2025).

Training uses a masked $y_{\text{prev}}$1-prediction objective. A binary mask $y_{\text{prev}}$2 is selected as a full mask $y_{\text{prev}}$3 of the time or one of six partial masks—anterior, posterior, inferior, superior, left, right—$y_{\text{prev}}$4 of the time. The context patch is constructed as

$y_{\text{prev}}$5

with $y_{\text{prev}}$6. The loss is

$y_{\text{prev}}$7

The paper does not report classifier-free guidance, external guidance scales, or DDIM inversion. Only the target patch is noised; the conditioning patch $y_{\text{prev}}$8 is not. Sampling starts from Gaussian noise, uses $y_{\text{prev}}$9 denoising steps, and carries previously refined context forward through $y_{\text{prev}}$0.

The coarse generator is itself specified. The autoencoder is a 3D ResNet-based encoder-decoder with two downsampling stages, corresponding to $y_{\text{prev}}$1 spatial downsampling, LeakyReLU activations, AdamW with learning rate $y_{\text{prev}}$2 and weight decay $y_{\text{prev}}$3, and output scaled to $y_{\text{prev}}$4. The latent diffusion backbone is a 3D U-Net with three scales and channels $y_{\text{prev}}$5, two residual blocks per level, self-attention at the last two scales, a $y_{\text{prev}}$6-channel latent input/output, a $y_{\text{prev}}$7-step linear $y_{\text{prev}}$8 schedule with $y_{\text{prev}}$9 and $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$0, velocity prediction, $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$1 training epochs, learning rate $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$2, batch size $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$3, and latent size $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$4. RealDeal itself uses a 3D U-Net with channels $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$5, two residual blocks per level, self-attention at the final scale, the same $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$6-step scaled linear $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$7 schedule, patch size $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$8, $$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I),$$9 training epochs, batch size $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$0, and base learning rate $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$1.

4. Data, evaluation protocol, and empirical findings

The reported experiments use Human Connectome Project T1-weighted structural MRI. The dataset contains $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$2 subjects, with an $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$3 train-test split, corresponding to $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$4 training subjects and $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$5 test subjects. Volumes are cropped to $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$6 voxels, retain the HCP isotropic voxel size of $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$7, and are intensity-rescaled to $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$8 using predefined global thresholds. No skull stripping or bias field correction is reported (Zhu et al., 24 Jul 2025).

Evaluation is deliberately broader than FID alone. The paper uses LPIPS with AlexNet patchwise across the three anatomical axes, FID with a patch-based protocol across three anatomical planes, coverage and density in feature space, KL divergence between empirical noise distributions estimated after ANTS-based denoising, sharpness via patch Laplacian variance after Gaussian smoothing $$\alpha_t = 1 - \beta_t, \qquad \bar{\alpha}_t = \prod_{s=1}^t \alpha_s,$$9 voxels, and HOG-based texture similarity over whole-brain and cerebellum masks.

For reconstructed-image experiments, RealDeal improves multiple realism indicators relative to coarse reconstructions. For fully synthetic-image experiments, it narrows the gap between LDM outputs and real data but does not reach the real-vs-real baseline.

Setting	Baseline	RealDeal
LPIPS, dim0/dim1/dim2	0.078 / 0.073 / 0.079	0.029 / 0.027 / 0.030
Noise KL divergence, white matter	0.775	0.149
Sharpness, Laplacian variance	0.0063	0.0111

For texture similarity, whole-brain HOG $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$0 distance changes from $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$1 to $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$2, and cerebellum from $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$3 to $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$4. Original-image sharpness is reported as $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$5, reconstructed-image sharpness as $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$6, and RealDeal-refined sharpness as $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$7. Additional noise KL divergence in the lateral ventricles changes from $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$8 to $$q(x_t \mid x_0) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\,x_0, (1-\bar{\alpha}_t)I).$$9.

For fully synthetic images, FID across the three anatomical planes changes from $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$0 for the LDM to $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$1 after RealDeal refinement, while the real-vs-real baseline is $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$2. Coverage at $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$3 changes from $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$4 to $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$5, and density from $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$6 to $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$7. Supplementary values at $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$8 and $$p_\theta(x_{t-1} \mid x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}),$$9 show the same directional pattern.

The paper also states several limitations. Patch-wise refinement can, in principle, introduce seam artifacts; mitigation is attempted through $\epsilon$0 conditioning and traversal order, but quantitative seam analysis is not reported. The forward process uses Gaussian noise rather than an explicit MRI-specific noise model such as Rician noise. Generalization is evaluated only on T1-weighted HCP data, without multi-site, multi-scanner, T2, or FLAIR validation. The paper further notes the risk of hallucinating details, implying that clinical deployment would require careful validation.

5. RealDeal-style negotiation as uncertainty-aware conversation forecasting

In dialogue research, RealDeal is not a named task in FortUne Dial. Instead, FortUne Dial formalizes “conversation forecasting” with uncertainty-aware metrics and then explains how the framework can be adapted to a RealDeal-style dyadic bargaining benchmark with outcomes such as deal/no-deal, price acceptance, and agreement success (Sicilia et al., 2024).

The formal setup treats a partial multi-party dialogue as $\epsilon$1, with a binary outcome $\epsilon$2, and a forecaster

$\epsilon$3

Calibration is expressed as

$\epsilon$4

and forecasting quality is evaluated with strictly proper scoring rules. The Brier score is

$\epsilon$5

with

$\epsilon$6

while the binary log score is

$\epsilon$7

A strong-reference comparative measure, Llama Skill Score, uses the Brier score of Llama-2-chat 70B direct forecasts as $\epsilon$8.

Two uncertainty representations are studied. Implicit forecasts use the next-token distribution under a yes/no prompt:

$\epsilon$9

Direct forecasts ask the model to verbalize a probability and then parse it:

$f_\psi$0

The paper proposes supervised fine-tuning with cross-entropy for IF, and off-policy reinforcement learning for DF with reward

$f_\psi$1

importance weighting, and PPO-style clipping. Both IF and DF can then be corrected with unified estimated-logit scaling:

$f_\psi$2

The negotiation corpora span eight tasks, including CraigslistBargain, CaSiNo, Persuasion for Good, and Deal or No Deal, as well as several non-bargaining conversational domains. Train/validation/test splits use original splits when available and otherwise $f_\psi$3. Each epoch samples approximately $f_\psi$4 dialogues per held-in dataset, with random truncation to simulate forecasting as a dialogue evolves. The paper also constructs easy, medium, and hard domain-generalization splits that hold out entire datasets.

For a RealDeal-style benchmark, the adaptation recommended in the paper is explicit: define $f_\psi$5 if a deal is reached, or if the agreed price is less than or equal to a target price; optionally define multiple one-vs-all outcomes; build dyadic bargaining splits with held-out categories, item types, or lengths; and evaluate with BS, BSS, LSS, log score, and bias. The same document further notes that, if in-domain validation data are available, IF with supervised fine-tuning and post-hoc scaling tends to yield the largest in-domain improvements, whereas DF with off-policy RL tends to generalize more consistently under zero-shot OOD deployment. It also reports that smaller open-source models can be calibrated to compete with pre-trained models $f_\psi$6 their size, and that a tuned $f_\psi$7B model can rival a $f_\psi$8 larger model in-domain.

6. RealDeal-style negotiation agents and the decoupling of strategy from generation

The phrase “RealDeal-style systems” is used most directly in work on negotiation agents that bargain over goods and prices while maintaining controllable strategy and human-like language. The central proposal is a modular architecture that separates high-level strategy from surface realization: a policy selects a coarse dialogue act such as propose(price=50) or counter(price=75), and a retrieval-based generator renders that act into contextually appropriate language (He et al., 2018).

The bargaining setting is defined around an initial asking price $f_\psi$9, side information such as title, description, and condition, and a dialogue that ends either with agreement at price $[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$0 or with no deal. The action inventory includes greet, inquire(info_type), inform(info_type, value), propose(price=x), counter(price=x), accept(price=x), reject, justify(reason), and thanks/closing. A policy $[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$1 selects actions given a state summarizing role, listing metadata, price history, the opponent’s last act or offer, and remaining time or turns. For acts requiring a price parameter, the model predicts either a classification over price bins or a regression target.

Surface generation is retrieval-based. Candidate human utterances are filtered by act type and parameter compatibility and then scored by

$[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$2

where $[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$3 includes TF-IDF overlap with listing text, act compatibility indicators, price compatibility features such as $[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$4, and local context similarity. For price-bearing acts, the generator may select templates with a placeholder and fill in the chosen value or prefer utterances whose original price is close to $[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$5 and optionally rewrite the number.

Training proceeds in two stages. Supervised learning minimizes act and price losses,

$[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$6

and, for regression,

$[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$7

with total objective

$[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$8

Reinforcement learning then fine-tunes only the act/price policy. The buyer reward is

$[x_t, \hat{x}_{\text{patch}}, y_{\text{prev}}]$9

and the seller reward is

$y_{\text{prev}}$00

optimized with REINFORCE:

$y_{\text{prev}}$01

The motivation for this decoupling is the failure mode of end-to-end token-level RL. When rewarded purely on utility, such systems can learn ungrammatical or repetitive outputs, invent odd codes, or adopt off-putting tactics. Freezing the generator while updating only the policy over acts and prices constrains exploration to an interpretable strategy space and preserves human-like language because utterances are retrieved from human corpora rather than optimized directly against reward.

The paper evaluates the approach on DEALORNODEAL and a richer Craigslist-based dataset. Its abstract states that human evaluation shows higher task success rate and more human-like negotiation behavior than previous approaches. Within the broader RealDeal-style framing, this modular line of work complements the FortUne Dial forecasting perspective: the former addresses controllable execution of bargaining strategy, while the latter addresses calibrated prediction of dialogue outcomes under uncertainty.