Colorimetric thermometry estimates the temperature of a hot object from a photograph. It compares how bright the object is in two or three color channels, and the ratio of those brightnesses fixes the temperature through Planck's law. The method is cheap and works with ordinary cameras. It fails, though, exactly where the picture is hardest to take. Flame brightness rises steeply with temperature, so in a single exposure the faint outer edge of a flame sinks into sensor noise while the bright core saturates. At the edge, the ratio of two noisy near-zero numbers is meaningless. Ignoring those weak boundary rings produces soot errors up to 80.97% (Edge Effects on Simultaneous Reconstructions of Flame Temperature and Soot Volume Fraction Profiles by a CCD Camera). The standard workaround, fusing eight exposures, cannot be used on turbulent or transient flames (Temperature Measurement Method of Flame Image Fusion with Different Exposures).
We propose a diffusion model, meaning a generator that learns to turn random noise into samples step by step, that works directly on a two-channel physical field: per-pixel emission amplitude and temperature. It is conditioned on one raw camera frame through an explicit sensor noise and clipping model. It does not denoise the RGB image. The output is a set of physically consistent temperature maps drawn from the posterior, from which we read uncertainty, "is there any emitter here at all" probabilities, and integrated quantities such as flame area or radiant power. Two design choices carry the proposal. The first replaces the learned image compressor with Planck coordinates, so the only information the representation discards is what is not graybody emission. The second uses a zero-terminal-SNR noise schedule, a noise schedule whose last step contains no trace of the clean signal. Our hypothesis is that the standard schedule's leaked dataset average shows up in thermometry as a measurable warm bias in dim and emission-free regions.
Core Claim
Common Diffusion Noise Schedules and Sample Steps are Flawed showed that common schedules leave a small fraction of signal at the final step. As a result, models drift toward mid-gray and cannot produce true black or true silence. They fix this with a zero terminal SNR schedule, velocity prediction (predicting a fixed mix of noise and clean signal, which stays stable when no signal remains), trailing timesteps, and rescaled guidance. They report COCO FID improving from 23.76 to 21.66.
Our claim is that the same flaw has a physical counterpart. In a model of temperature fields, "mid-gray" means "the average training temperature." Leaked-signal models should therefore fill noise-floor pixels with plausible warm emission rather than with no emission, which is the edge effect reproduced by a learned prior. We predict that the zero-SNR, velocity-prediction version removes this bias and produces calibrated uncertainty at the boundary. The comparison is against both a leaked-schedule twin and a direct regression network with the same backbone.
We do not claim more information than the frame contains. Theorem 5 of On The Classification-Distortion-Perception Tradeoff proves that processing a degraded observation cannot increase the information it carries about the target. Applied to temperature, this constrains us. A well-trained regressor already approximates the best point estimate, so any gain we report in mean temperature error at the edges must be small or suspect. Where the generative model can legitimately win is in representing the posterior correctly. At edge pixels the posterior is often bimodal, split between "no emitter" and "a weak emitter at some temperature." A per-pixel mean or a per-pixel Gaussian misrepresents that split. Field-level quantities, such as total radiant power or the isotherm area above a threshold, need spatially coherent joint samples, which per-pixel outputs cannot provide.
Research Program
Forward model. Each pixel is described by an amplitude $A$ (emissivity times optical thickness times geometry, all shared across channels under the graybody assumption) and a temperature $T$. The raw count in channel $c$ is modeled as
$$y_c = \mathrm{clip}\Big(\mathrm{Pois}\big(g\,\tau\,A\!\int S_c(\lambda)\,B(\lambda,T)\,d\lambda\big) + n_c,\ 0,\ y_{\max}\Big),$$
where $g$ is sensor gain, $\tau$ exposure time, $S_c$ the measured channel spectral response, $B$ the Planck function, $n_c$ Gaussian read noise with dark offset, and $y_{\max}$ the saturation level. Ratio pyrometry (Principles for a video fire detection system; Study of Tuyere Combustion Flame Temperature in Vanadium and Titanium Blast Furnaces by Machine Vision and Colorimetric Thermometry) is the noise-free inversion of this model. Our model keeps the noise and the clipping instead of dividing them away.
Physics as the bottleneck. In standard latent diffusion, an 8× learned compressor acts as a low-pass filter. Eliminating VAE for Fast and High-Resolution Generative Detail Restoration (GenDR-Pix) responded by removing the compressor. It packs pixels losslessly into channels (pixel-unshuffle) and adds a Masked Fourier Space loss against checkerboard artifacts. We reuse its pixel-unshuffle backbone and its Fourier loss, but we change what is being diffused. The model generates $(\mathrm{arcsinh}(A/a_0), T)$ fields rather than RGB, where $a_0$ is a small amplitude scale chosen so that zero emission is a finite, reachable value. Temperature is masked out of the loss where $A$ is below the photon floor. The model therefore learns "no emitter" as a distinct state instead of having to invent a temperature for empty pixels.
This representation is also an information bottleneck, but a physically meaningful one. Mapping three raw channels to two Planck coordinates throws away only the non-Planckian residual. We report that residual explicitly as a diagnostic for graybody violations, such as the unstable blue channel noted in the tuyere study.
Objective. The training loss is the velocity-prediction loss in physical coordinates, plus a data-consistency term at sampling time:
$$\mathcal{L} = \mathbb{E}\,\|v_\theta(x_t,t,y) - v_t\|^2 + \lambda_{\mathrm{F}}\,\mathcal{L}_{\mathrm{MFS}},\qquad \text{guidance: } \nabla_{x}\log p(y\mid A,T).$$
Here $x_t$ is the noised physical field at step $t$, $v_t$ the velocity target, $y$ the raw frame, $\mathcal{L}_{\mathrm{MFS}}$ the Masked Fourier Space penalty, and $p(y\mid A,T)$ the Poisson-Gaussian-clipped likelihood implied by the forward model.
This design responds to the loss-weight brittleness of composite objectives. Planck consistency holds by construction, so it needs no weight. The likelihood term is a proper probability rather than a perceptual heuristic, so the only weight left to tune is $\lambda_{\mathrm{F}}$. Clipped pixels contribute a one-sided likelihood, which tells the model only that the true value exceeds the saturation level. As a result, saturated cores should produce widened temperature intervals rather than confident wrong values.
Training data. We will render synthetic temperature and soot fields through the forward model across exposures and noise levels. The fields include the 17-ring axisymmetric coaxial flame of the edge-effect study, non-axisymmetric procedural turbulent fields, and line-of-sight projections of volumes reconstructed by Radiative Transport Based Flame Volume Reconstruction from Videos. We will then fine-tune on real calibrated frames.
Baselines
The naive reference is classical two-color ratio pyrometry applied to the single short-exposure frame, with the effective gray-level window of 50 to 220 used in the multi-exposure fusion study. Pixels outside that window are reported as invalid. This shows how much of the flame a practitioner simply loses. Where multiple exposures exist, we also run the eight-exposure fusion itself, which serves as the reference truth on real candle data.
The strongest classical method is regularized LSQR inversion over the full flame boundary, from Edge Effects on Simultaneous Reconstructions of Flame Temperature and Soot Volume Fraction Profiles by a CCD Camera. On its own axisymmetric benchmark at 46 dB SNR, it reports maximum average relative errors of 0.11% for temperature and 2.06% for soot volume fraction. We will run it on exactly that benchmark, where it is designed to excel, and on our non-axisymmetric fields, where its geometric assumption breaks. Beating it means matching its 0.11% on the axisymmetric case within noise and clearly outperforming it on boundary-ring error once axisymmetry is removed.
The strongest learned competitor, and the real threat, is a direct regression network. It uses the same pixel-unshuffle backbone, the same parameter count, the same synthetic and real data, and the same GPU-hour and hyperparameter-search budget. It is trained with a per-pixel heteroscedastic Gaussian likelihood on $(A,T)$ and outputs a mean and variance at each pixel. On high-temperature hardware snapshots the reference point is Dispersive Meta-lens Thermometry for High-temperature Measurements. It reports under 0.32% relative error on blackbodies from 1673 to 2973 K and under 1.5% on flame impingement, but it fails below about 1230 K. We cite it as the accuracy ceiling for well-exposed hot targets, not as a direct comparison, because it uses custom optics.
Three ablations isolate the key ingredients:
- An identical model trained with the standard leaked schedule and epsilon prediction, which tests the zero-SNR hypothesis.
- GenDR-Pix-style pixel-space diffusion on RGB, with an added Planck ratio-consistency loss and temperature computed afterward. This is the "obvious" combination and tests the physical-coordinates choice.
- The same RGB diffusion run through a frozen Stable Diffusion autoencoder, which tests the latent bottleneck.
All share data, backbone size, and compute budget.
Why not just…?
Why not just train a regression network with per-pixel uncertainty? This is the most serious alternative. By the information-loss principle above, its mean prediction should be nearly as good as ours. We expect it to fall short in two specific ways. First, a single Gaussian per pixel cannot express the bimodal emitter/no-emitter split at the boundary. Second, independent per-pixel errors give wrong uncertainty for field-level quantities such as flame area or total radiant power. Our experiments score coverage of 90% intervals for those integrated quantities and for detection of emission-free pixels. If the heteroscedastic regressor matches us on both, we will report that the generative machinery is unnecessary for this application.
Why not just use classical denoising plus LSQR or multi-exposure fusion? Multi-exposure fusion needs eight sequential shots spanning 20 to 200 µs, which turbulent flames and impact flashes lasting "0.1 s or less" (Low dispersion spectra of lunar impact flashes) do not allow. LSQR is strong but relies on concentric-ring geometry, while real tuyere flames show circumferential asymmetries of about 100 °C. Our synthetic suite contains both geometries, so this alternative is tested where it should win and where it should not, rather than dismissed.
Why not just do pixel-space RGB diffusion with a Planck ratio loss, or fine-tune Stable Diffusion? The first is a plausible competitor and is included as ablation (2). If it matches us, the physical-coordinates choice is not worth its complexity, and we will say so. For the second, we will measure the problem directly rather than assume it. We pass clean rendered flame frames through the SD autoencoder with no editing at all and measure the resulting temperature error per signal-level bin. If that round-trip error is already comparable to the target accuracy, latent diffusion is ruled out on data, not on argument. FLAME Diffuser: Wildfire Image Synthesis using Mask Guided Diffusion, which composites flames in pixel space before encoding with SD 1.5, is the natural instance to test this way.
Proposed Experiments
The first decisive check is the warm-bias test. On synthetic frames with known emission-free regions, we sweep the terminal signal fraction of the schedule from the standard value down to zero. For each setting we measure:
- the mean temperature assigned to true no-emission pixels;
- the false-emitter rate in those pixels;
- temperature bias in the lowest-photon-count decile of emitting pixels.
The hypothesis is confirmed if bias falls monotonically with the leaked signal fraction and disappears at zero SNR. It is refuted if the schedule makes no difference once the likelihood guidance term is present. That outcome would itself be informative: it would mean data consistency already corrects the leak.
The second check is the round-trip autoencoder diagnostic described above. It is cheap and determines whether the bottleneck argument holds for flames.
The main evaluation reports four kinds of results. The first is temperature relative error and soot or amplitude error, binned by photon count. The second is boundary-ring error on the coaxial benchmark. The third is the interval-score family: continuous ranked probability score and 90% interval coverage for per-pixel temperature, for flame area above an isotherm, and for total radiant power. The fourth is behavior in saturated cores: whether interval width grows with the clipped fraction.
Real-data tests use three sources:
- Blackbody furnace frames over 1073 to 1473 K, in the range of the candle study's calibration, at deliberately short exposures. Known temperature is the target.
- Single short-exposure candle frames, scored against the eight-exposure fusion reference.
- Tuyere frames, checked for whether the published roughly 100 to 120 °C asymmetry between tuyeres is preserved and whether the posterior widens rather than collapses in dust-occluded regions.
If access to the original tuyere and candle captures cannot be obtained, we will recapture equivalent blackbody and candle sequences with a calibrated RGB camera.
The proposal succeeds if zero-SNR physical-coordinate diffusion removes the warm bias, matches the regressor on point error, and clearly beats it on coverage of boundary and integrated quantities. It fails, and we will report it as failing, if the heteroscedastic regressor or the RGB-diffusion ablation reaches the same coverage.
References
- Edge Effects on Simultaneous Reconstructions of Flame Temperature and Soot Volume Fraction Profiles by a CCD Camera
- Temperature Measurement Method of Flame Image Fusion with Different Exposures
- Common Diffusion Noise Schedules and Sample Steps are Flawed
- On The Classification-Distortion-Perception Tradeoff
- Principles for a video fire detection system
- Study of Tuyere Combustion Flame Temperature in Vanadium and Titanium Blast Furnaces by Machine Vision and Colorimetric Thermometry
- Eliminating VAE for Fast and High-Resolution Generative Detail Restoration
- Radiative Transport Based Flame Volume Reconstruction from Videos
- Dispersive Meta-lens Thermometry for High-temperature Measurements
- Low dispersion spectra of lunar impact flashes
- FLAME Diffuser: Wildfire Image Synthesis using Mask Guided Diffusion