---
title: Temperature-Conditioned Diffusion Model
url: https://www.emergentmind.com/topics/temperature-conditioned-diffusion-model
type: topic
---

# Temperature-Conditioned Diffusion Model

A temperature-conditioned diffusion model is a diffusion-based generative or reconstruction model in which temperature is not merely an output variable but an explicit condition on the reverse process. In current arXiv usage, this usually denotes models that generate or reconstruct air temperature, land surface temperature, or climate fields from temperature observations, monthly means, or low-resolution temperature inputs, often together with auxiliary geophysical covariates such as LST, LULC, DEM, or sparse sensor masks [2412.13504] [2506.00033] [2511.05964]. A distinct but related usage treats “temperature” as an inverse-temperature or sampling-temperature control parameter that modulates diffusion dynamics, resampling, or reward tilting rather than conditioning on a physical temperature field [2604.10961] [2508.12361].

## 1. Definition and conceptual scope

In the meteorological, remote-sensing, and climate-emulation literature, a temperature-conditioned diffusion model is a conditional DDPM- or latent-diffusion-type model that learns a prior over temperature fields and then samples from a posterior induced by partial observations or physically relevant covariates. The conditioning signal may be sparse temperature observations and masks, land surface temperature, low-resolution temperature grids, monthly mean temperature maps, or coarse atmospheric predictors [2506.00033] [2412.13504] [2304.11699].

The conditioning target varies by application. In sparse interpolation, the objective is to reconstruct a full 2D temperature field from partial observations, formalized as sampling from $p(x \mid y, M)$ where $x$ is the full-state field and $M$ is an observation mask [2506.00033]. In urban air-temperature prediction, the target is $2\,\mathrm{m}$ air temperature $T_a$ at $100\,\mathrm{m}$ ground separation distance, generated from LST, LULC-derived features, and metadata [2412.13504]. In land-surface downscaling, the target is high-resolution LST $\mathbf{T}_{HR}$ conditioned on low-resolution LST $\mathbf{T}_{LR}$ and high-resolution geophysical priors $\mathbf{G}_{HR}$ [2511.05964]. In climate emulation, a spatio-temporal diffusion model is conditioned on monthly averages of temperature or precipitation on a $96 \times 96$ global grid and generates 28 daily fields that are realistic and consistent with those averages [2304.11699].

A plausible implication is that “temperature-conditioned” has become an umbrella term spanning at least three regimes: conditional reconstruction of temperature fields, conditional generation of temperature from related physical observables, and algorithmic control of diffusion dynamics via a temperature-like scalar. Because these regimes use the same diffusion vocabulary but different semantics, the term is inherently polysemous.

## 2. Probabilistic formulations and conditioning operators

A recurrent formulation is Bayesian posterior sampling. For sparse interpolation over the Southern Great Plains, the problem is posed as reconstructing a full $64 \times 64$ temperature field $x$ from sparse observations $y = M \odot x + \eta$, with conditional inference goal
$$
p(x \mid y, M) \propto p(y \mid x, M)\, p(x),
$$
where $p(x)$ is learned by a DDPM and conditioning is applied through mask-based inpainting and prekriging [2506.00033]. The reverse chain uses RePaint-style mask conditioning:
$$
x_{t-1} = M \odot x_{t-1}^{known} + (1-M) \odot x_{t-1}^{unknown},
$$
with
$$
x_{t-1}^{unknown} \sim \mathcal{N}(\mu_\theta(x_t,t), \Sigma_\theta(x_t,t)),
$$
and observed pixels injected at the correct diffusion noise level via
$$
x_{t-1}^{known} = \sqrt{\bar{\alpha}_t}\, y + \sqrt{1-\bar{\alpha}_t}\,\epsilon_t.
$$
The model augments the mask using ordinary kriging: a prekriged field $K$ is computed, pixels with kriging uncertainty below the $5$th percentile among unknowns are promoted to “known,” and mask-conditioned DDPM is then run with $(y',M')$ [2506.00033].

A second pattern is physically anchored interpolation between temperature variables. In DiffTemp, the standard terminal pure-noise state is replaced by the LST latent, yielding a deterministic LST-anchored schedule
$$
z_t = \sqrt{\bar{\alpha}_t}\, z_0 + \sqrt{1-\bar{\alpha}_t}\, z_y,
$$
with $z_0 = \mathrm{VAE.enc}(T_a)$ and $z_y = \mathrm{VAE.enc}(\mathrm{LST})$ [2412.13504]. The reverse model uses $v$-prediction, conditioned through ControlNet residuals derived from LST, LULC images, and metadata embeddings. The reverse update is
$$
z_{t-1} =
\sqrt{\bar{\alpha}_{t-1}}
\left[
\sqrt{\bar{\alpha}_t}\, z_t
-
\sqrt{1-\bar{\alpha}_t}\, v_\theta(z_t,t,c)
\right]
+
\sqrt{1-\bar{\alpha}_{t-1}}\, z_y.
$$
This schedule is explicitly intended to constrain generation to physically plausible $T_a$ near LST [2412.13504].

A third pattern is residual diffusion around a low-resolution temperature baseline. PGDM writes
$$
\mathbf{T}_{HR} = \mathbf{T}_{LR}^{\uparrow} + \Delta \mathbf{T},
$$
and formulates the task as
$$
\mathbf{T}_{HR} \sim p(\mathbf{T}_{HR}\mid \mathbf{T}_{LR}, \mathbf{G}_{HR}).
$$
Its ResShift forward process is centered on the residual $\Delta \mathbf{T}$ rather than on pure Gaussian corruption:
$$
q(\mathbf{x}_t \mid \mathbf{x}_{t-1}, \mathbf{T}_{LR})
=
\mathcal{N}\!\Big(
\mathbf{x}_t;\,
\mathbf{x}_{t-1}-\alpha_t \Delta \mathbf{T},
\, K^2 \alpha_t \mathbf{I}
\Big),
$$
with reverse mean parameterized through a denoised estimate
$$
\widehat{\mathbf{x}_0}
=
\mathbf{T}_{LR}^{\uparrow} + f_\theta(\mathbf{x}_t,t,\mathbf{c}).
$$
The training loss is direct MSE on $\mathbf{x}_0$ rather than $\epsilon$- or $v$-prediction [2511.05964].

DiffESM uses a different conditional structure. It is a continuous-time diffusion model with $v$-parameterization, conditioned on a spatial map of monthly mean temperature or precipitation, the day of the year that the 28-day sequence begins on, and the diffusion time [2304.11699]. The paper does not report an explicit aggregation penalty enforcing exact monthly mean consistency; consistency is learned implicitly from conditioning. This distinguishes it from methods that directly enforce pixelwise or coarse-grid reconstruction constraints.

## 3. Architectural realizations

The literature contains several distinct architectural realizations, but all retain the basic diffusion decomposition into a learned prior and a conditional reverse operator.

| System | Conditioning signal | Backbone |
|---|---|---|
| Sparse interpolation | Sparse observations, binary mask, prekriged field | OpenAI guided-diffusion UNet [2506.00033] |
| DiffTemp | LST, RGB, NDVI, NDBI, NDWI, metadata | Stable Diffusion VAE + U-Net + ControlNet [2412.13504] |
| PGDM | LR LST, reflectances, NDVI/NDWI/NDMI, DEM, LULC | Dual-branch UNet-like encoder–decoder [2511.05964] |
| PDE-informed downscaling | ERA5 predictors, static high-resolution maps | Residual latent diffusion model with VAE and UNet-like denoiser [2510.23866] |
| DiffESM | Monthly mean map and day-of-year | Fully convolutional spatio-temporal U-Net [2304.11699] |

In sparse interpolation, the DDPM uses OpenAI’s guided-diffusion UNet backbone, predicts $\epsilon$ rather than $x_0$ or a score, learns $\Sigma_\theta$ with `learn_sigma=True`, and is trained with $T=250$ diffusion steps, multi-resolution attention at $64 \times 64$, $32 \times 32$, $16 \times 16$, and $8 \times 8$, Adam with learning rate $3 \times 10^{-4}$, batch size $1$, and $100\mathrm{k}$ epochs for $64 \times 64$ inputs [2506.00033]. Inpainting uses timestep re-spacing to $150$ steps and RePaint resampling with $r=10$, $j=10$.

DiffTemp is a latent diffusion model built on Stable Diffusion with ControlNet conditioning. The denoising U-Net is pretrained on satellite imagery via DiffusionSat and fine-tuned here; ControlNet ingests conditioning images and injects residuals at multiple scales. The optimizer is Adam with learning rate $5\times 10^{-5}$, batch size $4$, and training lasts $300\mathrm{k}$ steps total, with fine-tuning $3\mathrm{k}$ steps per city [2412.13504].

PGDM adopts a dual-branch encoder–decoder. One branch is state-aware and processes $(\mathbf{x}_t,t)$ through time-adaptive ResBlocks and Multi-Head Non-Local attention; the second branch processes $(\mathbf{T}_{LR}^{\uparrow}, \mathbf{G}_{HR})$. The decoder uses PixelShuffle upsampling and U-Net skip connections, with a final residual head predicting $\Delta \mathbf{T}$ [2511.05964]. The selected configuration uses base width $C_{\text{base}}=32$ because it gives RMSE $0.610$ with $12.17\mathrm{M}$ parameters and $57.86\mathrm{G}$ FLOPs, while $C_{\text{base}}=64$ improves RMSE to $0.574$ but raises complexity to $47.72\mathrm{M}$ parameters and $230.55\mathrm{G}$ FLOPs [2511.05964].

The PDE-informed downscaling model uses a VAE to encode high-resolution temperature, a pre-trained reference UNet upscaler, and a latent diffusion denoiser that learns the residual relative to the upscaler output [2510.23866]. Because of memory limits, only the final $\sim 200$ million parameters are updated during fine-tuning. DiffESM, by contrast, is fully convolutional in $C \times T \times H \times W$ format, with interleaved temporal and spatial convolution layers, no self-attention, four downsampling/upsampling levels, and per-level channel widths $48$, $128$, $192$, $256$ [2304.11699].

## 4. Physics guidance and physically informed conditioning

The major technical divergence among temperature-conditioned diffusion models concerns whether physics enters only through conditioning variables or through the loss itself.

DiffTemp relies on an LST-anchored latent schedule and on physically interpretable covariates rather than on an explicit physics loss. LST is treated as a “physical boundary,” while RGB, NDVI, NDBI, NDWI, and metadata encode urban fabric, seasonality, and geolocation [2412.13504]. The paper states that no explicit clipping is described; instead, anchoring the terminal state at LST and conditioning on LULC plus metadata act as physical priors. This design also enables counterfactual simulation: edited RGB/LULC can be passed through a separate RGB $\rightarrow$ LST diffusion model, and the modified LST and RGB are then fed to DiffTemp to estimate $\Delta T_a = \hat{T}_a(c') - \hat{T}_a(c)$ under altered urban layouts [2412.13504].

PGDM is more explicit. It grounds its conditioning variables in the surface energy balance
$$
R_n = H + LE + G,
$$
and treats $\mathbf{p}_{HR}$, $\mathbf{NDXI}_{HR}$, $\mathbf{H}_{HR}$, and $\mathbf{LULC}_{HR}$ as proxies for albedo, emissivity, vegetation cover, aerodynamic resistance, surface resistance, and related SEB terms [2511.05964]. Although the analytic SEB residual is not directly evaluated during training, the model monitors physical consistency using an energy-conservation degradation operator $D(\cdot)$ and
$$
\mathcal{L}_{\text{phy}}
=
\frac{1}{N}\sum_{k=1}^{N}\frac{1}{M}
\left\|
D\!\big(\widehat{\mathbf{T}_{HR}^{(k)}}\big)-\mathbf{T}_{LR}^{(k)}
\right\|_2^2.
$$
The paper emphasizes that this term is used in evaluation rather than training [2511.05964].

The PDE-informed latent diffusion model goes further by incorporating a physics-informed loss directly into training:
$$
L_{\mathrm{total}} = L_{\mathrm{diffusion}} + \lambda_{\mathrm{PDE}} L_{\mathrm{PDE}}.
$$
Its physical constraint is not a full prognostic PDE but a diagnostic flux-ratio consistency between coarse and generated fine fields, derived from an effective advection–diffusion balance and computed in decoded pixel space [2510.23866]. The reported effect is specific: conventional residual LDM training already yields small PDE residuals, while the added loss further reduces the flux-ratio discrepancy and improves spectral consistency, though $LDM_{PDE}$ “typically has slightly higher RMSE and lower $R^2$ than the best purely statistical baselines” [2510.23866].

DiffESM occupies the opposite end of the spectrum. It conditions on monthly mean maps and day-of-year labels but does not introduce an explicit consistency loss
$$
L_{\mathrm{cons}} = \lambda \|A(x_{1:D}) - M\|_F^2,
$$
which the paper presents only as a reference form not used in training [2304.11699]. This suggests a broader methodological split: some temperature-conditioned diffusion models encode physics via priors and covariates, whereas others regularize the reverse process or decoded field by explicit diagnostic constraints.

## 5. Applications, empirical behavior, and uncertainty

The application space spans sparse-data assimilation, urban microclimate mapping, LST downscaling, atmospheric downscaling, and climate emulation.

For sparse interpolation, four methods are compared: base diffusion, KrigSCD, inverse distance weighting, and conditional Gaussian simulations. Evaluation covers $1\%$, $5\%$, $10\%$, $20\%$, and $30\%$ known coverage [2506.00033]. KrigSCD achieves the lowest LPIPS across all coverages, with average LPIPS $0.165$ versus $0.186$ for base diffusion, $0.327$ for IDW, and $0.282$ for CGS. By $20\%$ known coverage, reconstructions are visually close to ground truth, with LPIPS $\approx 0.047$ for KrigSCD. Pixelwise errors behave differently: at $1\%$ known, RMSE is $\approx 32.20$ for IDW, $\approx 36.12$ for CGS, $\approx 101.42$ for base diffusion, and $\approx 83.05$ for KrigSCD, while by $30\%$ known KrigSCD reaches RMSE $\approx 11.56$ [2506.00033]. The paper’s interpretation is explicit: classical methods minimize RMSE/MAE by design at very low coverage, but diffusion models better recover spatial patterns, urban cold spots, gradients, and texture.

DiffTemp reports same-resolution performance on LSTAT-20K with RMSE $2.20$, MAE $2.09$, and SSIM $0.70$, compared with Random Forest at RMSE $3.72$, MAE $3.45$, SSIM $0.32$; Gradient Boosting at RMSE $3.44$, MAE $3.25$, SSIM $0.36$; Linear Regression at RMSE $2.84$, MAE $2.63$, SSIM $0.23$; and MLP at RMSE $3.54$, MAE $3.42$, SSIM $0.45$ [2412.13504]. In super-resolution, the model achieves RMSE $0.920$, MAE $0.849$, SSIM $0.779$ when downsampled $T_a$ at $300\,\mathrm{m}$ is used as an extra condition, and Point SR yields RMSE $1.707$, MAE $1.588$, SSIM $0.724$ when $N$ points from $T_a$ emulate station measurements [2412.13504]. The noise-schedule ablation is especially decisive: a pure-noise terminal schedule yields RMSE $6.09$, MAE $6.26$, SSIM $0.44$, whereas the LST-anchored schedule yields RMSE $2.20$, MAE $2.09$, SSIM $0.70$ [2412.13504].

PGDM reports the strongest benchmark detail. On Landsat_CN20 test data, it attains RMSE $0.610\,\mathrm{K}$ and SSIM $0.973$ at $10\times$, and RMSE $0.749\,\mathrm{K}$ and SSIM $0.967$ at $20\times$; the corresponding $\mathcal{L}_{\text{phy}}$ values are $0.025\,\mathrm{K}$ and $0.042\,\mathrm{K}$ [2511.05964]. On Landsat_GLB it reaches RMSE $0.910/1.126$ and SSIM $0.949/0.937$, and on ASTER_GLB RMSE $1.764/1.930$ and SSIM $0.849/0.833$, outperforming bilinear interpolation, kernel-driven methods, DCF, and MoCoLSK-Net across the reported datasets [2511.05964]. Its stochasticity is also exploited for self-assessment: with $S=5$ samples, the scene-level mean diffusion standard deviation $\overline{\sigma}_{\text{diff}}$ has a strong positive linear correlation with actual scene-level MAE,
$$
\mathrm{MAE} \approx 2.95\,\overline{\sigma}_{\text{diff}} - 0.03,\qquad r=0.971.
$$

The PDE-informed downscaling model evaluates both statistical and physics-aware scores. It reports that $LDM_{PDE}$ achieves the best physics-aware scores, specifically the lowest $L_{\mathrm{flux}}$ and the lowest median and tightest interquartile range in $L_{\mathrm{spec}}$, while visual comparisons show preserved fine filamentary structures and reduced speckle relative to the base residual LDM [2510.23866]. DiffESM, finally, is assessed through spatial maps and histograms of Monthly Hot Streak, Monthly Hot Days, 90th Quantile Values, Monthly Dry Spell, Monthly Dry Days, and SDII; generated differences resemble validation-test differences, indicating that the emulator reproduces the ESM’s spatio-temporal distributional characteristics, with closer agreement for temperature than for precipitation and slight under-prediction bias in precipitation [2304.11699].

## 6. Ambiguities, limitations, and related meanings of “temperature”

A common misconception is that “temperature-conditioned diffusion model” always means a generative model conditioned on a temperature field. In several lines of work, “temperature” is instead a control parameter governing sampling sharpness or the effective dynamics of the reverse process.

In the statistical-mechanics analysis of discrete diffusion models, the forward kernel can be written in a Boltzmannized form with effective inverse temperature
$$
F_t := \tanh^{-1}(\theta^t),
$$
which controls coupling between $x_t$ and $x_0$ [2604.10961]. The paper identifies a speciation transition at
$$
F_t^2 \Lambda = 1
$$
and a collapse transition at the REM threshold $s_t=0$. This temperature is not meteorological; it is an effective inverse temperature induced by the noise schedule.

In inference-time scaling with SMC, temperature again denotes reward sharpness:
$$
p_{\mathrm{tar}}(x,c) = \frac{1}{Z} p_{\mathrm{pre}}(x,c)\exp(\lambda r(x,c)),
$$
with adaptive temperature
$$
\lambda_t = \frac{T-t}{T}\lambda
$$
used to down-weight unreliable early rewards and emphasize later, more reliable ones [2508.12361]. This is orthogonal to classifier-free guidance and unrelated to physical temperature fields.

There is also an older physical-transport usage in which diffusion itself is conditioned by thermodynamic temperature. Within the Haken–Strobl–Reineker SQLE, the long-time diffusion coefficient obeys
$$
D_\infty(T)=\frac{c_1}{T}
$$
for diagonal dynamical disorder, and
$$
D_\infty(T)=\frac{c_1}{T}+c_2 T
$$
for diagonal plus off-diagonal disorder under white-noise, classical-bath assumptions [2406.12750]. Here the model is not a generative diffusion model at all; it is a transport theory in which temperature modulates dephasing rates and therefore the diffusion coefficient.

Across the generative literature, the main limitations are explicit. Sparse interpolation in the SGP domain uses a discretized $0$–$255$ pixel range and no explicit physical constraints, which the paper identifies as a limitation for operational meteorology [2506.00033]. DiffTemp notes possible limitations during extreme heat waves and unusual wind regimes because wind and humidity are absent from conditioning, and further validation with in-situ data is needed [2412.13504]. The PDE-informed downscaler approximates advection direction from the temperature gradient rather than explicit winds in the loss term and is evaluated only over Italy [2510.23866]. PGDM omits explicit atmospheric and radiation fields, using DEM and LR LST as proxies, and its evaluation follows an upscaling–downscaling protocol [2511.05964]. DiffESM does not enforce exact monthly-mean consistency, models temperature and precipitation separately, and reports only one ESM family and a limited scenario set [2304.11699].

Taken together, these works indicate that the phrase “temperature-conditioned diffusion model” should be interpreted contextually. In contemporary geoscientific machine learning it usually denotes a conditional generative model for temperature fields, often strengthened by geophysical priors or physics-aware losses. In discrete diffusion theory and inference-time scaling, it instead refers to a temperature-like scalar that governs diversity, fidelity, or phase transitions in the generative process itself.

Source: https://www.emergentmind.com/topics/temperature-conditioned-diffusion-model