- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies multiple appropriate facial reaction generation, where several plausible listener facial reactions are generated from a speaker's audio-visual behaviour. Existing diffusion methods denoise directly from Gaussian noise to the listener reaction, which the authors argue starts too far from the target. Because listeners tend to mimic speakers, the paper instead diffuses the difference between the listener and speaker emotion representations. At inference the residual is added back to the speaker representation and passed through a fixed feasibility projection. The authors also standardize each residual dimension with training-set statistics so that all dimensions share the same signal-to-noise ratio at each timestep. Experiments on the MARS dataset with the REACT 2025 splits compare against Trans-VAE, REGNN and PerFRDiff, and include a controlled ablation of the diffusion target and an analysis of the denoising trajectory.
Critical strengths
The controlled ablation is clean. Table 2 changes only the diffusion target while keeping the architecture, noise schedule and conditioning fixed, and FRCorr rises from 0.69 to 0.80, a gain that can be attributed fairly clearly to the residual target. The trajectory analysis in Table 4 also shows that residual diffusion already reaches at its first step the level the baseline needs about sixty percent of the denoising steps to reach, which is consistent with the authors' motivation
Critical weaknesses
The first problem is limited novelty. Diffusing the residual between a conditioning anchor and the target is well established in image restoration and generation, for example ResShift and ResDiff for super-resolution, as well as RDDM, InDI, I2SB and bridge diffusion models in general. None of this work is cited or discussed, and the claim of being the first to formulate the task this way holds only within the narrow scope of facial reaction generation. Per-dimension standardization is routine preprocessing. Overall the paper applies an existing technique to a specific task, and the methodological contribution for the ICLR audience is weak.
Second, the results do not convincingly show that the method is better. The improvement is almost entirely in FRCorr. On FRDist the two variants score 180.7 and 199.0, both worse than PerFRDiff at 177.8 and clearly worse than REGNN at 141.9, and the normalized variant has an FRDiv of only 0.097 compared with 0.139 for PerFRDiff. More importantly, the B_Mime baseline, which simply copies the speaker, already reaches an FRCorr of 0.52, close to the 0.56 of PerFRDiff. FRCorr thus rewards outputs that stay close to the speaker, and it is hard to tell whether the model learns appropriate reactions or small perturbations of the speaker's expressions. In addition, only one dataset is used, no variance over runs is reported, there is no comparison with recent diffusion methods such as ReactDiff, SC-Diff, VQ-Diff and SAFRG, and there are no rendered results or user studies. Three different models in Tables 2 and 3 have exactly the same FRSyn of 45.73, which the authors should check.
Questions for authors
- How far are the generated reactions from the speaker representation, and how does that compare with the distance between real listeners and the speaker? If the baseline is the speaker representation plus the mean residual or plus random noise, does ResDiffFRG win on all metrics?
- Why do the normalized variant, its version without CFG and the Stochastic Baseline have identical FRSyn values?
- Could you report the mean and standard deviation over several random seeds, and compare with ReactDiff, SC-Diff and VQ-Diff under the REACT 2025 protocol?
- Apart from the choice of anchor, what is the essential technical difference between this work and ResShift, RDDM, InDI or bridge diffusion?