A diffusion model generates an image by starting from random noise and denoising it over many steps. To run on a phone, it is usually quantized: its weights and activations are stored as 4- or 8-bit integers instead of 32-bit floats. The resulting rounding error is set once, offline, against a full-precision reference and a fixed calibration set. On the device, nothing checks whether that calibration still holds after the user changes the sampler, the number of steps, or the kind of prompt. We propose to borrow the central idea behind Cronbach's alpha from psychometrics: the reliability of a measurement can be estimated without knowing the true value, by checking how well several parallel measurements of the same quantity agree. A quantized layer can be made to produce two such parallel measurements cheaply. We run the same activation through the integer grid twice, each time with a different small random offset added before rounding (a technique called dithering), so that the two rounding errors are independent. The agreement between the two outputs then estimates the layer's signal-to-quantization-noise ratio without any full-precision teacher.
This estimate gives an edge runtime something it currently lacks: a live, unit-free gauge of how much of each layer's activation is signal and how much is rounding noise, available at every denoising step. We use it in two ways. First, it detects when deployed conditions have moved outside the calibration regime. Second, it drives a controller that raises precision, from 4-bit to 8-bit activations, only at the steps and layers where measured reliability falls below a threshold, so that a fixed average bit budget is spent where it matters for the current run.
Core Claim
In classical test theory, an observed score equals a true score plus independent error, and reliability is the share of observed variance that is true-score variance. Write $\rho = \sigma_T^2/(\sigma_T^2+\sigma_E^2)$, where $\sigma_T^2$ is the variance of the full-precision activation across the tokens of a layer and $\sigma_E^2$ is the variance of the quantization error. If two dithered forms $a$ and $b$ have independent errors of equal variance, their correlation across tokens is an unbiased-in-the-limit estimate of $\rho$. The quantity engineers actually care about then follows directly:
$$\widehat{\mathrm{SQNR}} = \frac{r_{ab}}{1-r_{ab}},$$
where $r_{ab}$ is the across-token correlation between the two forms. This identity comes from $\rho/(1-\rho) = \sigma_T^2/\sigma_E^2$. With $K>2$ forms, Cronbach's alpha computed over forms generalizes the same estimate. Alpha's equivalence to the mean of all split-half coefficients, which the resilience scale study uses in its psychometric setting, is what justifies treating the two-form correlation as a special case.
The claim is narrow and testable. The reference-free estimate should track the true, teacher-based SQNR per layer and per step closely enough to rank where quantization hurts. That ranking should be good enough to detect calibration shift and to allocate precision better than fixed schedules, at an overhead of a few percent of runtime.
We are deliberately not proposing what the research file's mapping table suggests, namely computing alpha across channels or attention heads as a coherence score. Channels in a layer are not meant to measure one shared quantity. High alpha across them would mostly indicate redundancy, and nothing in the file shows such a score predicts image quality. The psychometric literature itself warns against reading one aggregate alpha as structural health. In the German Short Stress State Questionnaire validation, a total alpha of .85 coexisted with a six-factor structure that differed from the original three factors. We therefore borrow the parallel-forms logic of reliability, which is sound here, and not alpha-as-unidimensionality, which is not.
Research Program
The work builds on the timestep-aware quantization line. Q-Diffusion calibrates activation ranges on latents sampled across the whole denoising trajectory. PTQD splits quantization noise into a part correlated with the layer output and an uncorrelated part, then corrects both and fixes a step-aware 4/8-bit mixed-precision schedule. TAC-Diffusion stores per-timestep bias and scale correction tables. Meta-weighted PTQ learns calibration sample and frequency weights by bi-level optimization. All four make their decisions offline against a full-precision model. The file identifies their shared weakness: correction tables and step sizes degrade when the sampler, step count, or prompt distribution changes at deployment. Our work supplies the missing on-device measurement and does not replace these methods; each one can be used as the base quantizer.
PTQD's analysis also constrains the design. Deterministic rounding produces error that is partly correlated with the signal. Two deterministic forms would share that error, and their agreement would overstate reliability. Dithering is what breaks this. Subtractive dither makes rounding error approximately independent of the input and of the other form. Verifying this empirically is the first work package. We will measure the cross-form error correlation per layer with a full-precision reference. We will also check the psychometric "parallel forms" assumption of equal error variance across forms, borrowing the logic of the invariance testing that the TNAS longitudinal study applies across measurement waves. Here the "waves" are denoising steps and the question is whether the estimator's assumptions hold uniformly across them.
The second work package sets the cost. Dithering only activations means one copy of the weights, so the second form costs a second integer matmul on the monitored subset only. We will monitor a random 2–5% of tokens, a few layers per U-Net stage, and every $m$-th step. We will then report the sampling error of $r_{ab}$ as a function of token count, using the standard Fisher-z interval, to pick the smallest subsample that ranks layers stably. Weight-quantization error is invisible to this estimator because both forms share the same quantized weights. We state that scope plainly: the method monitors activation quantization, which is the binding constraint at W4A4.
The third work package is the controller. At each monitored step, layers whose estimated SQNR falls below a threshold calibrated offline switch to 8-bit activations for the next window of steps. A global budget caps the average activation bit-width. The fourth work package is a shift detector: an alarm fires when the reliability profile across steps departs from the profile recorded at calibration time. Only that profile, a few hundred scalars, is stored; no teacher model is kept on the device.
Baselines
The naive reference point is the quantized model with no monitoring and a uniform bit-width, run with each base quantizer's published calibration. It is what ships today, and every comparison is reported against it at the same weight precision.
The strongest existing competitor for precision allocation is PTQD's step-aware mixed precision, which spends 8 bits on late, noise-sensitive steps and 4 bits on early ones. It is the strongest because it already exploits the known timestep structure and comes with a reported ImageNet LDM-4 FID of 6.44, 1.39 above full precision, and FID about 7.75 at 20 steps. For a fair comparison, our controller runs on top of the same PTQD-corrected base model and calibration set, with the same average activation bit-width and the same sampler. It beats PTQD only if FID is lower at matched bits and matched measured latency, or equal FID is reached at fewer bits. A gain that appears only once our monitoring overhead is excluded does not count.
For the ultra-low-bit regime, the reference is TAC-Diffusion. The file reports an FID of 9.55 at W3A8 on CIFAR-10, against 17.31 for Q-Diffusion, and an FID of 7.79 at W2A8 on LSUN-Bedrooms. We layer monitoring on top of TAC and test whether escalation recovers quality under sampler shift, where TAC's fixed tables are expected to degrade.
The oracle upper bound is the same controller driven by the true, teacher-based SQNR. It shows how much of the achievable gain the reference-free estimate captures. Two further references isolate the key ingredient. The first is a random-escalation controller at the same budget. The second is a teacher-free heuristic controller driven by activation range (maximum absolute value), the cheap signal a runtime might use instead.
The decisive ablation compares deterministic two-form agreement with dithered agreement. If deterministic forms work equally well, independence of errors is not doing the work and the psychometric justification collapses.
Why not just…?
Why not just measure SQNR against the full-precision model? That is the oracle. On a phone, keeping the FP16 teacher defeats the memory savings that motivated quantization. SnapFusion notes that parameter footprint already strains lower-end devices even after step reduction. The oracle comparison above measures exactly how much is lost by going reference-free. If the gap is large, the method fails on its own terms.
Why not just use a fixed mixed-precision schedule like PTQD's? This is the real threat. If the reliability profile is nearly identical across samplers, step counts, and prompts, a fixed schedule captures most of the benefit and per-run monitoring adds only overhead. The calibration-shift experiments settle this. We expect fixed schedules to lose ground when calibration on 100-step DDIM is deployed with a 20-step solver. PTQD's own report that Q-Diffusion collapses to FID above 100 at 20 steps shows how sharply step count can break calibration. If they do not lose ground, we will report that adaptive allocation is unnecessary, and the contribution reduces to the shift detector.
Why not just use activation range or simple disagreement norms? Range statistics measure scale, not noise share. A wide range with a fine enough grid is harmless. The raw mean-squared disagreement between dithered forms equals twice the error variance and is not normalized by signal variance, so one threshold cannot transfer across layers whose activations differ by up to 200x, as Q-Diffusion reports for shortcut layers. The reliability ratio is unit-free, which is the property a single global threshold requires. Both heuristics are run as baselines.
Why not just distill to fewer steps? Step distillation, as in SnapFusion's reduction from 50 to 8 steps, cuts the number of passes but not the per-pass bit-width. Its training cost is large. It is complementary, and we include an 8-step regime to check that monitoring still helps when there are few steps over which to amortize its cost.
Proposed Experiments
We use the benchmarks the quantization papers in the file share. These are CIFAR-10 at 32×32 and LSUN-Bedrooms and LSUN-Churches at 256×256 for unconditional models, ImageNet 256×256 with LDM-4 for class-conditional generation, and Stable Diffusion on the MS-COCO 2017 5K subset for text-to-image, scored with FID and CLIP score. The file reports no result on Lacuna for monitored or adaptive-at-runtime precision in any of these settings. All published numbers above serve only as anchors for our reimplementations, and we rerun every baseline in our own pipeline before comparing.
The first decisive check is estimator validity. Across W4A4 and W4A8, all monitored layers, and all steps, the rank correlation between estimated and oracle SQNR should exceed 0.9. Cross-form error correlation should remain near zero under dither and clearly positive without it. If the estimate does not track the oracle, we stop.
The second check is shift detection. We calibrate on DDIM at 100 steps with one prompt set, then deploy DPM-Solver at 20 steps and an out-of-distribution prompt set. The question is whether the profile-deviation alarm separates degraded runs from undegraded ones, with degradation defined by oracle per-sample error. We report the area under the detection curve.
The third check is matched-budget allocation. The controller is compared against PTQD's schedule, against TAC with and without monitoring, and against the oracle, random, and range-based controllers at equal average activation bits, both in-distribution and under shift.
The fourth check is cost. We measure end-to-end latency for the U-Net on an RTX 3090 with integer kernels, the platform where PTQD reports a 3.34x speedup at 4 bits, and on a mobile neural processor. The monitoring overhead must stay below 5% for the method to be worth deploying.
References
- The resilience scale: factorial structure, reliability, validity, and parenting-related factors among disaster-exposed adolescents
- The Short Stress State Questionnaire in German (SSSQ-G)
- Q-Diffusion: Quantizing Diffusion Models
- PTQD: Accurate Post-Training Quantization for Diffusion Models
- Timestep-Aware Correction for Quantized Diffusion Models
- Beyond Uniformity: Sample and Frequency Meta Weighting for Post-Training Quantization of Diffusion Models
- Psychometric properties and longitudinal measurement invariance of the treatment-induced neuropathy assessment scale in the Chinese cancer chemotherapy population
- SnapFusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds