Paper · 2025-12-22 · stat.ML, cs.LG, stat.ME
Diffusion Models in Simulation-Based Inference: A Tutorial Review
Tutorial review of diffusion and flow-matching models for simulation-based inference: training objectives, SDE/ODE samplers, design choices (schedules, weighting, parameterisation, solvers), compositional scores for pooled data. Benchmarks: VP-EDM diffusion and flow matching are most reliable; consistency models are fastest but weaker at high dimension. Author-reported.
Diffusion modelsSimulation-based inferenceFlow matchingNoise schedulediffusion-modelssimulation-based-inferencetutorialflow-matchingcompositional-inference
arXivPDFCodeMarkdown16 min read
AI-compiled by claude-haiku-5-5 on 2026-11-13, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
In plain words
Scientists often have a simulator, a program that produces fake data from a set of hidden settings, but no formula for how likely a given dataset is under given settings. Simulation-based inference (SBI) asks the reverse question: given real data, which settings could have produced it? The approach is to run the simulator many times and train a learned model on the pairs of settings and results.
Diffusion models are a family of generative models that learn to turn random noise into a plausible sample. Here they are trained to produce the hidden settings, conditioned on the observed data. The same machinery can also produce the data given the settings, or both together. The paper is a guide to the many ways this can be built and tuned, plus a set of experiments that compare the choices on the same problems.
The main lessons are practical. Several choices, such as how noise is scheduled over time, how the training loss is weighted, what the network is asked to predict, and which solver draws the samples, change both accuracy and speed. Some faster model families trade accuracy for speed. Combining the learned models for many datasets, or steering them with extra information at sampling time, is also possible without retraining.
TL;DR
A tutorial review of diffusion models for SBI, with four benchmark studies. Across the low-dimensional benchmark, a variance-preserving EDM diffusion model with F-prediction and flow matching were the most reliable choices, and an adaptive two-step SDE solver was the most accurate sampler §6.1. Consistency models are much faster at inference but less accurate, and they lose more ground as dimension grows §6.2, §6.3. The paper's own post-publication note says the loss weights were not applied per sample in the reported runs, so weighting results need care §6.
Problem
Simulation-based inference targets the posterior over parameters given observations when the likelihood cannot be evaluated but simulations can be drawn §2.1. Training data come from the joint model; a generative network is then fitted so that it can be queried for the posterior (neural posterior estimation), the likelihood (neural likelihood estimation), or the joint §2.1, §3.
The cost of these methods is set by how many simulations are available, how many posteriors must be computed, and how high the dimensions of parameters and observations are §2.2. The paper defines efficiency as total time over simulation, training and inference, and accuracy through a proper distance between the true and learned posterior. Because that distance cannot be computed exactly, proxies such as simulation-based calibration, the classifier two-sample test and posterior predictive checks are used, each with known limits §2.2.
Method
Forward and reverse process §3. Clean targets z_0 are corrupted to z_t = alpha_t z_0 + sigma_t eps with Gaussian eps. The schedule (alpha_t, sigma_t) defines the noise level over time; the paper's central design variable is this schedule §5.1, §5.2. Diffusion is written as an SDE, and sampling runs the reverse SDE or the matching probability-flow ODE from a Gaussian base distribution at t = 1 down to t = 0 §3, §5.5.
Training objective §3. The score of the noised marginal is estimated by denoising score matching: a network regresses onto the conditional score, which is available in closed form. Equivalent forms predict the noise, the clean target, a velocity, or the EDM-reconditioned target. These forms are the same loss up to a noise-level-dependent weight §5.4. A conditional version uses the condition x as an extra input §3.
Flow matching §3, §5.6.1. With alpha_t = 1 - t and sigma_t = t, a network regresses onto the difference between noise and clean data along the path. The learned vector field defines the same probability flow as the ODE above. A power-law distribution over t (rho = -0.6 in the experiments) changes which noise levels get weight §5.6.1.
Likelihood evaluation §3. Log densities come from integrating the probability-flow ODE while tracking the divergence of the velocity field. The trace of the Jacobian is estimated stochastically with the Hutchinson estimator §3.
Samplers §5.5. SDE solvers include Euler–Maruyama and higher-order additive-noise methods (SEA, ShARK), and an adaptive two-step method that the paper finds best in its experiments §5.5, §6.1. ODE solvers include Euler and adaptive Runge–Kutta methods app. A.1.1. Annealed Langevin dynamics and predictor–corrector sampling are also described; the paper found no accuracy gain from them and discourages their use for SBI §5.5, §6.1.
Guidance and inference-time control §4.1. Classifier-free guidance estimates the likelihood score as the difference between the conditional and unconditional score, so prior and likelihood can be scaled separately §4.1.1. Soft constraints can be added as a log-sigmoid penalty term scaled by a function of the noise level §4.1.1. Simulator feedback can be used during training: a primary flow is trained with self-conditioning, then a small control flow corrects it at small diffusion times §4.1.2. Prior changes can be applied at sampling time with a correction derived from the prior ratio §4.1.3. The paper warns that any guidance biases the reverse process, so posterior calibration must be checked §4.1.1.
Compositional scores §4.2. Because a product of likelihoods becomes a sum of scores, a model trained on single observations or single groups can be combined for many. The compositional identity for N observations weights the prior score by (1 - N)(1 - t) and sums the single-observation scores §4.2.1. It is exact only at t = 0, so error accumulates, and the paper uses a damping schedule and mini-batches over groups to control it §4.2.1, §4.2.3. Partial pooling is handled with a global model and a local model, sampled ancestrally §4.2.3.
Structured targets §4.3. Multiscale generation uses coarse-to-fine factorisation; causal structure can be built into the network; a model trained on the joint distribution with random masks can answer any conditional query §4.3.1–§4.3.3. Scores on manifolds and infinite-dimensional spaces are discussed §4.3.4.
Consistency models §5.6.2. A consistency function maps any noisy state at any time directly to the clean target, with a skip-connection parameterisation that fixes the boundary condition at t = 0. Sampling takes one step from the base distribution, or a few refinement steps (10 for the discrete model and 15 for the continuous model in the experiments) §5.6.2, app. A.1. Two training routes are described: distilling a trained diffusion model, and training as a standalone estimator. The SBI experiments use the standalone route. A limitation is that density evaluation at arbitrary points is not available §5.6.2.
Pseudocode for the training step (noise parametrisation). Equivalent parametrisations and weights are in §5.3, §5.4.
# Denoising training step, noise parametrisation [§3] [§5.4]
# Input: targets (B, D); conditions (B, C)
# Output: scalar loss
def training_step(targets, conditions, network):
# Draw one noise level per example, uniform on (0, 1) [§3]
times = sample_uniform(0, 1, size=(B,))
# Draw the Gaussian noise that corrupts each target [§3]
noise = sample_normal(size=(B, D))
# Forward process z_t = alpha_t * z_0 + sigma_t * noise [§3]
noisy = alpha(times) * targets + sigma(times) * noise
# The network predicts the noise from the noisy target and condition [§5.4]
predicted = network(noisy, conditions, times)
# Weighted squared error in noise space; omega is a design choice [§5.3]
loss = omega(times) * squared_error(predicted, noise)
return mean(loss)Pseudocode for posterior sampling with an Euler solver on the probability-flow ODE. SDE samplers replace the drift step with a stochastic step §5.5.
# Posterior draws by Euler steps on the probability-flow ODE [§3] [§5.5]
# Input: condition (C,); number of draws N; number of steps K
# Output: draws (N, D)
def sample_posterior(condition, network, num_draws, steps):
# Start from the base distribution at t = 1 [§3]
states = sample_normal(size=(num_draws, D))
# Build a decreasing time grid from 1 down to 0 [§5.5]
times = linspace(1, 0, steps + 1)
for time, next_time in zip(times[:-1], times[1:]):
# Evaluate the learned drift at this state and time [§5.5]
drift = probability_flow_drift(network, states, condition, time)
# Euler step; next_time < time, so the step moves toward t = 0 [§5.5]
states = states + drift * (next_time - time)
return statesPseudocode for the compositional score over R groups in the complete-pooling case. The stabilised form with damping and mini-batches is in §4.2.3.
# Compositional score for complete pooling over R groups [§4.2]
# Input: state (D,); time t; prior score; per-group scores
# Output: approximate score of the pooled posterior at state
def pooled_score(state, time, prior_score, group_score, damping):
# The prior enters with weight (1 - R) and factor (1 - t) [§4.2]
prior_term = (1 - R) * (1 - time) * prior_score(state)
# Random mini-batch of M group indices at each reverse step [§4.2.3]
chosen = sample_groups(R, M)
# Scale the mini-batch sum up to all R groups [§4.2.3]
group_term = (R / M) * sum(group_score(state, time, g) for g in chosen)
# Damping d(t) controls error accumulation [§4.2.3]
return damping(time) * (prior_term + group_term)Hyperparameters
Values are from the appendix and the experiment text unless noted. Where the paper says "not stated", the repository was checked and is noted in Paper vs code.
| Name | Value | Where | Notes |
|---|---|---|---|
| Case study 1 budget | 1,000 epochs; 32,000 new simulations per epoch; batch 128 | app. A.1.1 model_settings_benchmark.py | Text says 30,000 new simulations per epoch §6.1; see Paper vs code |
| Optimiser and learning rate | Adam (online) or AdamW (offline); cosine schedule; initial 5e-4 | app. A.1 | Case study 3 uses AdamW with 1e-4 app. A.1.3 |
| Backbone (MLP) | 5 layers, width 256, Mish, He-normal init, residual connections | app. A.1 model_settings_benchmark.py | Dropout 0.05 for all models app. A.1 |
| Time embedding | Fourier features, dimension 32; FiLM on the time embedding; layer norm after skip | app. A.1 | |
| Transformer backbone | 5 blocks, width 128, 4 heads, SwiGLU (expansion 3), QK RMS norm, AdaLN-Zero | app. A.1.1 | About 3.5x slower in training and inference §6.1 |
| EDM noise schedule | rho = 7; sigma_max = 80; sigma_min = 0.002 in text, 0.0001 in appendix; sigma_target = 1 | §5.2 app. A.1 | The two sigma_min values disagree; see Pitfalls |
| Cosine schedule | Variance-preserving; shift s = 0; lambda in [-15, 15] | app. A.1 | Weighting differs between the text and the code; see Paper vs code |
| EDM weighting | omega = 1 + exp(lambda) / sigma_target^2 | §5.3 | Applied with the EDM schedule in the EDM runs app. A.1 |
| Flow-matching power law | rho = -0.6 for time sampling | §5.6.1 §6.1 | Helped on low-dimensional tasks only §5.6.1 |
| Optimal transport | Entropic regularisation 0.1; 100 Sinkhorn iterations; partial mass 0.8 (OT-partial) or 0.9 (combined); condition ratio 0.01 for COT | app. A.1.1 | Batch-level coupling |
| EMA | gamma = 6.94, beta_i = (1 - 1/i)^(gamma + 1) | app. A.1 ema_callback.py | No systematic gain reported §6.1 |
| Sampler (case study 1) | 500 steps; adaptive methods bounded to 50–1000 steps | app. A.1.1 model_settings_benchmark.py | Adaptive two-step SDE: absolute tolerance 0.02576, relative tolerance 0.1 app. A.1.1 |
| Consistency sampling steps | 10 (discrete), 15 (continuous) | app. A.1 | Not verified in the repository (see Paper vs code) |
| Case study 2 budget | 32,768 simulations, fixed; 1,000 test datasets; 10 training runs | §6.2 app. A.1.2 | Batch size and epochs not stated in the paper |
| Case study 2 training (code) | Batch 64; 1,000 epochs | model_settings.py | Taken from the code only |
| Case study 2 networks | Residual MLP backbone (5 × 256); summary: FusionTransformer or LSTNet; summary dimension twice the parameter count | app. A.1.2 | Classifier for C2ST: two MLPs of width 128, 5-fold CV |
| Case study 3 priors | log sigma ~ N(0, 0.3²); alpha ~ N(3, 0.5²); spectrum P(k) ∝ k^(-alpha) sigma² | app. A.1.3 §6.3 | Unit length proportional to 1/ |
| Case study 3 resolutions | 8, 16, 32, 64, 128, 256 (256 only for the low-dimensional target) | §6.3 app. A.1.3 | |
| Case study 3 training | AdamW, learning rate 1e-4, batch 32; 5,000 training fields; 500 epochs | app. A.1.3 §6.3 | Offline: 5,000 fixed simulations; online: 5,000 new per epoch, 2.5M in total |
| Case study 3 sampling | TSIT5 ODE, 500 steps (DM, FM); 10 steps (CM) | app. A.1.3 | |
| Case study 4 model | EDM VP training; 500 epochs; 256 batches of 128; R = 1 subject per training instance; N = 30 trials | app. A.1.4 settings.py | SetTransformer summary dimension 16, dropout 0.1 |
| Case study 4 inference | 100 test datasets of 100 subjects × 30 trials; two-step adaptive SDE, max 1000 steps | app. A.1.4 | |
| Case study 4 damping | d(t) = d0 (d1/d0)^t; d1 = 1; d0 = 0.1 (partial pooling), 0.01 (complete pooling) | app. A.1.4 complete_pooling.py partial_pooling.py | Code matches the appendix |
| Case study 4 mini-batch | M = 3 groups per reverse step | app. A.1.4 complete_pooling.py | |
| Case study 4 simulator | Euler–Maruyama step 1e-3; max decision time 10 s; start at beta · alpha; absorbing boundaries at 0 and alpha | app. A.1.4 | Priors listed in app. A.1.4 |
Results
Author-reported. The paper states the setup for each study; the runs are the ones the post-publication note describes, so the weighting issue under Pitfalls applies to them.
Case study 1: low-dimensional benchmarks §6.1, fig. 11. Ten benchmark problems from Lueckmann et al. (2021), scored with the classifier two-sample test (C2ST; 0.5 is best) on 10 datasets. Diffusion with a VP EDM schedule was the strongest overall, with cosine schedules a near second; VE EDM was worse on several tasks §6.1. Flow matching was competitive, with a slight drop on the smallest problems §6.1. Consistency models were close but slightly less accurate and much faster §6.1. The text gives no summary numbers for this study; use fig. 11 for values.
Case study 2: large ODE model §6.2, tab. 4. An ODE model of 72 parameters and about 27,000 measurements. Each configuration was trained in 10 repeated runs and tested on 1,000 datasets. MCMC was run on one dataset. Full results from tab. 4 (mean (standard deviation) over 10 runs; rank is by the sum of standardised metrics):
| Group | Variant | Solver | NRMSE | Miscalib. | Contraction | C2ST | TARP | Place |
|---|---|---|---|---|---|---|---|---|
| Diffusion | EDM, VP, F | ODE | 0.42 (0.005) | 0.03 (0.001) | 0.68 (0.004) | 0.57 (0.008) | -0.02 (0.005) | 1 (2) |
| Diffusion | Cosine, VP, v | SDE | 0.39 (0.005) | 0.04 (0.003) | 0.70 (0.003) | 0.59 (0.016) | -0.02 (0.004) | 3 (2) |
| Diffusion | Cosine, VP, v | ODE | 0.40 (0.005) | 0.04 (0.003) | 0.70 (0.004) | 0.58 (0.025) | -0.02 (0.004) | 4 (4) |
| Diffusion | EDM, VP, F | SDE | 0.40 (0.005) | 0.04 (0.002) | 0.70 (0.004) | 0.57 (0.015) | -0.03 (0.004) | 5 (3) |
| Diffusion | Cosine, VP, F | ODE | 0.40 (0.003) | 0.04 (0.002) | 0.70 (0.002) | 0.60 (0.016) | -0.01 (0.003) | 6 (3) |
| Diffusion | Cosine, VP, F | SDE | 0.39 (0.003) | 0.04 (0.001) | 0.71 (0.002) | 0.60 (0.017) | -0.02 (0.003) | 9 (3) |
| Diffusion | Cosine, VP, epsilon | SDE | 0.52 (0.129) | 0.04 (0.001) | 0.62 (0.122) | 0.61 (0.020) | -0.05 (0.006) | 12 (2) |
| Diffusion | EDM, VE, F | SDE | 0.67 (0.024) | 0.10 (0.003) | 0.51 (0.028) | 0.63 (0.013) | 0.15 (0.005) | 16 (1) |
| Diffusion | EDM, VE, F | ODE | 0.39 (0.015) | 0.15 (0.002) | 0.84 (0.006) | 0.70 (0.010) | -0.31 (0.012) | 17 (2) |
| Diffusion | Cosine, VP, epsilon | ODE | 1.38 (0.124) | 0.07 (0.005) | 0.01 (0.006) | 0.52 (0.005) | 0.12 (0.006) | 18 (1) |
| Diffusion (EMA) | EDM, VP, F | ODE | 0.42 (0.006) | 0.03 (0.002) | 0.68 (0.005) | 0.57 (0.010) | -0.03 (0.006) | 2 (2) |
| Diffusion (EMA) | EDM, VP, F | SDE | 0.40 (0.005) | 0.04 (0.003) | 0.70 (0.004) | 0.59 (0.018) | -0.04 (0.003) | 8 (4) |
| Flow matching | rho = -0.6 | ODE | 0.41 (0.005) | 0.05 (0.004) | 0.69 (0.005) | 0.59 (0.016) | 0.01 (0.005) | 7 (2) |
| Flow matching | Uniform | ODE | 0.39 (0.003) | 0.07 (0.002) | 0.77 (0.005) | 0.59 (0.016) | -0.06 (0.006) | 10 (2) |
| Flow matching | OT | ODE | 0.39 (0.003) | 0.07 (0.002) | 0.76 (0.004) | 0.59 (0.010) | -0.05 (0.004) | 11 (1) |
| Consistency | Discrete | ODE | 0.53 (0.006) | 0.11 (0.067) | 0.64 (0.010) | 0.58 (0.072) | 0.05 (0.051) | 13 (2) |
| Consistency | Continuous | ODE | 0.70 (0.010) | 0.14 (0.004) | 0.49 (0.013) | 0.52 (0.010) | 0.15 (0.003) | 15 (1) |
| MCMC | - | - | 0.18 (0.004) | 0.29 (0.002) | 1.00 (0.000) | 0.73 (0.026) | -0.09 (0.012) | 14 (0) |
Source: §6.2, tab. 4. Bold marks the best value in each column, as in the paper. Training of all 1,000 datasets took under a day, against about three days for MCMC on one dataset §6.2.
Case study 3: Gaussian random fields §6.3, fig. 13. Low-dimensional targets (alpha, log sigma from one field): NRMSE is similar across the three families; calibration favours diffusion at high resolution §6.3. High-dimensional targets (the field itself): diffusion and flow matching stay close to chance-level C2ST across scales and gain from online training; consistency models degrade with resolution §6.3. The paper reports that consistency models sample orders of magnitude faster in this setting §6.3.
Case study 4: pooling regimes §6.4, fig. 14, fig. 15. An evidence accumulation model with four parameters per trial. Partial pooling gives the lowest RMSE through shrinkage; complete pooling sharpens the posterior with more subjects; no pooling gives broad but calibrated posteriors §6.4. Larger numbers of groups need adaptive steps, and the damping factor trades calibration against sharpness; the mini-batch size has little effect §6.4, fig. 15.
Paper vs code
The code is at github.com/bayesflow-org/diffusion-experiments (commit 1765b807), owned by the BayesFlow organisation whose developers are authors; the paper links it in the appendix app. A.1. The repository has no LICENSE file, so no snippets are quoted here.
- Post-publication weighting. The paper's own boxed note says the weights acted as one rescaling per batch rather than per sample, so the reported diffusion models were effectively unweighted. It adds that the sigmoid variant did not work as designed, and that the continuous consistency model used a modified objective §6. The repository README repeats this and says BayesFlow 2.1.0 corrects both points. The weighting code lives in BayesFlow, which is not in this repository, so this was not checked here README.md.
- Case study 1 budget. Text: 30,000 new simulations per epoch §6.1. Appendix: 32,000 simulations per epoch app. A.1.1. Code: 32,000 // 128 = 250 batches of 128 per epoch, i.e. 32,000 model_settings_benchmark.py. The code agrees with the appendix.
- EDM sigma_min. Section 5.2 gives sigma_min = 0.002 §5.2; the appendix gives 0.0001 app. A.1. The repository sets only
noise_schedule: "edm"and does not pass sigma_min, so it uses the BayesFlow default, which is not visible here model_settings_benchmark.py. - Cosine weighting. The appendix says the cosine schedule uses sigmoid weighting app. A.1. In the code, only
diffusion_cosine_v_lwsetslikelihood_weightingexplicitly; the other cosine variants take the BayesFlow default model_settings_benchmark.py. The default is not visible in this repository. - BayesFlow version. The appendix says BayesFlow 2.0.12 app. A.1. The code pins
bayesflow==2.0.13pyproject.toml. - Partial-pooling local model. The text calls it a continuous consistency model §6.4. The code uses
bf.networks.StableConsistencyModelpartial_pooling.py. These may be the same family, but the paper does not name the class. - Case study 2 batch and epochs. Not stated in the paper. The code uses batch 64 and 1,000 epochs model_settings.py.
- Case study 3 learning rate. The code uses 1e-4 train_models_field_inference.py, which matches the appendix for that study app. A.1.3.
- Case study 4 damping and mini-batch. The code matches the appendix: d0 = 0.01 for complete pooling and 0.1 for partial pooling, and a mini-batch of 3 complete_pooling.py partial_pooling.py.
- Consistency sampling steps. The appendix gives 10 and 15 steps app. A.1. The compatibility rules in the code allow consistency models with the 500-step Euler sampler and the ODE variants model_settings_benchmark.py. The 10- and 15-step settings were not located in the code.
- Checked and matching: EDM and cosine parameter names, five-layer width-256 MLP, EMA gamma 6.94, sampler step counts (500, min 50, max 1000), and the case study 4 settings.
Pitfalls
- Weights were not applied per sample in the reported runs. The paper's own note says so §6. Any result that depends on EDM or sigmoid weighting should be re-run on BayesFlow 2.1.0 or later, as the repository recommends. Sigmoid weighting is marked as not recommended §6.
- sigma_min disagrees. Section 5.2 says 0.002 and the appendix says 0.0001 for the EDM settings §5.2 app. A.1. The code uses the library default. Check the value before implementing.
- Noise prediction is unstable here. Plain epsilon-prediction with Euler steps was prone to instability and produced divergent trajectories for about 5% of samples in case study 2 §6.2. The paper recommends v- or F-parametrisation §5.4, §7.
- Score prediction without weighting is equivalent to noise prediction with weight proportional to 1 / sigma_t^2 §5.4. Switching the parametrisation without changing the loss weight changes the effective training distribution.
- Compositional scores are exact only at t = 0 §4.2.1. Error accumulates with the number of groups. The damping d(t) and the adaptive solver are used to control it §4.2.3, §6.4.
- Guidance biases the reverse process §4.1.1. Check calibration after any guidance, for example with SBC or TARP §7.
- Langevin and predictor–corrector sampling did not improve accuracy and were harder to tune §6.1. The paper discourages them for SBI §6.1.
- VE schedules were worse than VP on the benchmarks, and in case study 2 VE showed higher calibration error §5.1, §6.1, §6.2.
- C2ST is fragile in high dimension. In case study 2 it stayed near chance even where other metrics showed clear differences §6.2. The paper calls it useful but fragile §7.
- Flow-matching OT cost. Section 5.6.1 says OT flow matching trains about 2.5x longer §5.6.1; case study 2 reports 1.5x §6.2. The two numbers come from different setups.
- Consistency models cannot evaluate density at arbitrary points §5.6.2. They are fastest but least accurate at high dimension §6.2, §6.3.
- Equation and figure numbers are derived from the LaTeX source, not from a compiled PDF. Equations are cited by section rather than number. Appendix numbers (A.1, A.1.1–A.1.4) follow the source order and were not checked against a rendered PDF.
- Figure-only results. Case studies 1, 3 and 4 give no summary numbers in the text. Values must be read from the figures.
Lineage
- Builds on: score-based SDE diffusion §3, §5.1; denoising score matching; the DDPM-style discrete sampler §3; conditional flow matching and rectified flow §5.6.1; the EDM design §5.2–§5.4; consistency models §5.6.2; and compositional score matching for SBI (Geffner et al., 2023) §4.2.
- Applied in SBI: the 2024–2025 works listed in the paper's summary table tab. 3. These include neural posterior estimation with flow matching (Wildberger et al., 2023), sequential score estimation (Sharrock et al., 2024), the joint-training approach (Gloeckler et al., 2024), prior guidance (Yang et al., 2025), and hierarchical compositional estimation (Arruda et al., 2025).
- Compared against: BayesFlow benchmarks (Lueckmann et al., 2021) in case study 1, and MCMC via PEtab and pyPESTO in case study 2 §6.1, §6.2.
- Superseded by: not stated. The paper calls the choice of loss weights and parametrisation open questions §7.
Source and license
Adapted from “Diffusion Models in Simulation-Based Inference: A Tutorial Review” by Jonas Arruda, Niels Bracher, Ullrich Köthe, Jan Hasenauer, Stefan T. Radev (arXiv:2512.20685), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors. Code citations refer to https://github.com/bayesflow-org/diffusion-experiments at commit 1765b807dbc2 (none (no LICENSE file in the repository)). Any quoted code is shown under that license.