--- title: Ambient-Consistent Distillation (AC-DMD) url: https://www.emergentmind.com/topics/ambient-consistent-distribution-matching-distillation-ac-dmd type: topic --- # Ambient-Consistent Distillation (AC-DMD) Searching arXiv for the cited paper and closely related distribution-matching distillation work. Ambient-Consistent Distribution Matching Distillation (AC-DMD) is the first-stage cold-start procedure introduced within Reward-Tilted Distribution Matching Distillation (RTDMD) for few-step flow generators. It matches a \(K\)-step student generator to a pretrained teacher on a sequence of subintervals \([t_k,1]\), while jointly training an auxiliary fake score model to approximate the student’s marginal scores on those same subintervals. Its distinguishing feature is a consistency regularizer applied to the fake score model, intended to help that model remain stable as the generator distribution shifts under limited updates. In the reported formulation, AC-DMD supplies a low-variance, scheduler-agnostic initialization for downstream reward-tilted optimization rather than serving as the complete alignment procedure by itself [2605.26108]. ## 1. Role within few-step reward-tilted distillation AC-DMD appears as the first of two stages in RTDMD. The full framework is described as unifying distribution matching distillation with reward-guided reinforcement learning for few-step flow generators. Within that decomposition, AC-DMD is the distribution-matching stage: it trains the student generator \(G_\theta\) against a teacher distribution \(p_\psi\) and simultaneously trains a fake score model \(s_\phi\) that estimates the student’s marginal scores over re-noised student latents. The second RTDMD stage then jointly optimizes the distribution-matching and reward-maximization terms; AC-DMD is the precursor that supplies the cold-start needed for that later optimization [2605.26108]. The setup is defined on latent variables. Let \(p_\psi\) denote the pretrained teacher distribution on latents \(x_0\), and let \(G_\theta\) denote the \(K\)-step student. After step \(k\), the student latent \(x_{t_k}\) has distribution \(p_\theta^{(k)}\). Rather than matching only a single terminal distribution, AC-DMD re-noises \(x_{t_k}\) forward to any \(t \in [t_k,1]\) and matches the resulting student marginal to the teacher over that subinterval. This subinterval-wise construction is the central structural difference emphasized in the method description. A plausible implication is that AC-DMD addresses a specific instability of few-step distillation: the score estimator used inside the generator update is trained on samples from a student distribution that changes during optimization. The consistency term is introduced precisely to regularize this estimator against such temporal drift. ## 2. Objective formulation and subinterval-wise matching The generator objective at step \(k\) is a reverse-KL over the subinterval \([t_k,1]\): $$ L_{\mathrm{gen}}^{(k)}(\theta) = \mathbb{E}_{t\sim \mathrm{Unif}[t_k,1]} \left[ \lambda_k(t)\cdot \mathrm{KL}\!\left(p_{\theta,t}^{(k)} \,\|\, p_{\psi,t}\right) \right]. $$ The corresponding pathwise gradient is approximated as $$ \nabla_\theta L_{\mathrm{gen}}^{(k)} \approx - \mathbb{E}_{t, x_{t_k}, \epsilon} \left[ \lambda_k(t)\,\alpha_k(t)\, \bigl(s_\psi(x_t^{(k)},t)-s_\phi(x_t^{(k)},t)\bigr)^\top \partial_\theta x_{t_k} \right], $$ where $$ \alpha_k(t)=\frac{1-t}{1-t_k}, \qquad \sigma_k(t)=\frac{t-t_k}{1-t_k}, $$ and $$ x_t^{(k)}=\alpha_k(t)x_{t_k}+\sigma_k(t)\epsilon. $$ The induced student marginal on the subinterval is $$ p_{\theta,t}^{(k)}(x) = \int q_t^{(k)}(x\mid x_{t_k})\,p_\theta^{(k)}(x_{t_k})\,dx_{t_k}, $$ with Gaussian forward kernel $$ q_t^{(k)}(x\mid x_{t_k}) = \mathcal{N}\!\left(x;\alpha_k(t)x_{t_k},\sigma_k(t)^2 I\right). $$ The fake score model is trained with an ambient denoising score-matching objective: $$ L_{\mathrm{fake}}^{(k)}(\phi) = \mathbb{E}_{t\sim U[t_k,1],\,x_{t_k},\,\epsilon} \left[ \omega_k(t)\cdot \left\| s_\phi(x_t^{(k)},t) - \nabla_x \log q_t^{(k)}(x_t^{(k)}\mid x_{t_k}) \right\|^2 \right], $$ where $$ \nabla_x \log q = \frac{\alpha_k(t)x_{t_k}-x}{\sigma_k(t)^2}. $$ In words, the generator update matches the student’s noised-latent distributions to the teacher on each subinterval, while the fake score model learns the student’s marginal scores on the same region. The summary identifies the DSM distance metric as \(L2\) on velocity/score difference, equivalent to KL [2605.26108]. ## 3. Forward re-noising construction and derivation assumptions The subinterval KL term is derived by starting from the student latent after \(k\) steps, \(x_{t_k}\sim p_\theta^{(k)}\), and forwarding it to a later noise level \(t\in[t_k,1]\) by $$ x_t^{(k)}=\alpha_k(t)x_{t_k}+\sigma_k(t)\epsilon, \qquad \epsilon\sim\mathcal{N}(0,I). $$ This yields the student marginal \(p_{\theta,t}^{(k)}\) used inside the reverse-KL objective. The generator gradient then takes the usual reverse-KL form involving the difference between the student marginal score \(\nabla_x \log p_{\theta,t}^{(k)}(x_t^{(k)})\) and the teacher score \(s_\psi(x_t^{(k)},t)\), after which \(\nabla_x \log p_{\theta,t}^{(k)}\) is approximated by the learned fake score \(s_\phi(x_t^{(k)},t)\) [2605.26108]. Three assumptions are stated explicitly for this derivation. First, the noise schedule is rectified-flow, with \(\alpha_t=1-t\) and \(\sigma_t=t\). Second, the sampling kernel from \(x_{t_k}\) to \(x_t\) is Gaussian with known \(\alpha_k\) and \(\sigma_k\). Third, the generator’s current mapping \(G_\theta\) induces \(p_\theta^{(k)}\) exactly via coefficient-preserving sampling (CPS). This framing distinguishes AC-DMD from full-interval formulations that match only over \([0,1]\). The paper summary states that base A-DMD generalizes DMD2 when \(\eta=1\) and the full interval \([0,1]\) is used, while remaining unbiased under any CPS scheduler. It also reports no quality loss when specializing to \(\eta=1\), but extension to \(\eta<1\). A common source of confusion is therefore to treat AC-DMD as merely a restatement of full-interval DMD2; the stated relation is instead one of generalization under CPS, with subinterval-wise matching as the operative construction. ## 4. Consistency regularization and fake-score stabilization The consistency term is introduced because the ambient DSM loss \(L_{\mathrm{fake}}\) is described as unbiased but high-variance when \(s_\phi\) is trained on fresh student latents that shift each iteration. The formulation uses the denoiser or “\(x\)-prediction” form \(f_\phi(\cdot,t)\) of the fake score model and imposes a self-consistency relation across adjacent noise levels: $$ L_{\mathrm{cons}}^{(k)}(\phi) = \mathbb{E}_{t',t'',\,x_{t_k}\sim p_\theta^{(k)},\,x_{t'}\sim q_{t'}^{(k)}(\cdot\mid x_{t_k})} \left[ \left\| f_\phi(x_{t'},t') - \mathbb{E}_{\tilde{x}_{t''}\sim p_\phi(x_{t''}\mid x_{t'})} \bigl[f_\phi(\tilde{x}_{t''},t'')\bigr] \right\|^2 \right]. $$ The optimal denoiser is stated to satisfy $$ f^*(x_{t'},t') = \mathbb{E}_{\tilde{x}_{t''}\sim p_\phi(\cdot\mid x_{t'})} \bigl[f^*(\tilde{x}_{t''},t'')\bigr] \qquad \text{for all } t''t_1>\dots>t_K=0\}\), weights \(\lambda_k(t)\), \(\omega_k(t)\), consistency weight \(\gamma\), and the number of cold-start steps \(T_{\mathrm{cold}}\). For each update \(u=1,\dots,T_{\mathrm{cold}}\), the sequence is: 1. Sample a mini-batch of prompts, or pure Gaussian noise for the unconditional case. 2. Roll out the \(K\)-step generator \(G_\theta\), recording each intermediate \(x_{t_k}\). 3. Uniformly sample a step index \(k\in\{1,\dots,K\}\) and a noise level \(t\sim U[t_k,1]\). 4. Form \(x_t^{(k)}=\alpha_k(t)x_{t_k}+\sigma_k(t)\epsilon\), with \(\epsilon\sim\mathcal{N}\). 5. Update \(\phi\) by descending \(\nabla_\phi L_{\mathrm{fake\_total}}^{(k)}(\phi)\). 6. Update \(\theta\) by descending \(\nabla_\theta L_{\mathrm{gen}}^{(k)}(\theta)\). The notes accompanying the pseudocode state that subintervals \([t_k,1]\) are inherent in the \(K\)-step schedule, that \(L_{\mathrm{cons}}\) is applied at every fake-score update, and that \(\lambda_k(t)\) and \(\omega_k(t)\) may be chosen constant or may follow the teacher’s weighting \(w(t)\). The hyperparameter and architecture choices are specified as follows. Typical \(K=4\), with $$ t_k=\{1,0.75,0.5,0.25,0\}. $$ The weights \(\lambda_k(t)\) and \(\omega_k(t)\) are typically uniform in \(t\) or inherited from the teacher’s \(w(t)=\alpha_t\). The consistency weight \(\gamma\) is ablated over \(\{0.001,0.005,0.01,0.1\}\), with best performance at \(\gamma=0.01\). The CPS stochasticity parameter \(\eta\) controls injection noise; \(\eta\ge 0.8\) works well, with default \(\eta=0.9\). Learning rates are reported as \(3\times 10^{-5}\) for SD3* and \(1\times 10^{-6}\) for FLUX.2 4B during the cold-start stage, shared by \(\theta\) and \(\phi\); during the RL stage, \(\phi\) retains the cold-start learning rate and \(\theta\) is reduced to \(3\times 10^{-6}\). The generator \(G_\theta\) uses the same UNet/flow backbone as the teacher and is finetuned via LoRA with rank \(64\) and \(\alpha=32\), while the fake score model \(s_\phi\) copies the teacher’s score-network architecture [2605.26108]. ## 6. Guarantees, ablations, and empirical role The theoretical guarantees summarized for AC-DMD are limited but concrete. Proposition 1, located in Appendix C, states that the ambient DSM objective \(L_{\mathrm{fake}}^{(k)}\) is unbiased and that its unique minimizer is the true marginal score \(\nabla_x \log p_{\theta,t}^{(k)}\). Proposition 2, in Appendix D, states that the optimal denoiser satisfies the self-consistency property used to justify \(L_{\mathrm{cons}}\). No formal global convergence proof is provided. The paper summary nevertheless reports that AC-DMD cold-starts reliably in practice and produces low-variance score estimates that accelerate downstream RL [2605.26108]. The ablations attribute a measurable share of performance to the consistency regularizer alone, before any RL is applied. For the consistency study in Table 6, base A-DMD with \(\gamma=0\) yields CLIPScore \(=0.2753\), PickScore \(=22.01\), and HPSv2 \(=0.3188\). AC-DMD with \(\gamma=0.01\) yields CLIPScore \(=0.2843\), PickScore \(=23.57\), and HPSv2 \(=0.3456\), corresponding to gains of \(+0.0090\), \(+1.56\), and \(+0.0268\), respectively. Improvement is reported as robust across \(\gamma\in[0.005,0.1]\). A separate ablation on the second RTDMD stage reports that naïve GRPO without consistency reaches PickScore \(\approx 23.30\), whereas SubGRPO with \(M=2\) plus deterministic final-step gradient reaches PickScore \(=23.57\) and CLIPScore \(=0.2843\), with gains of \(+0.26\) and \(+0.0084\). Each component, including shared-noise and final-step backpropagation, is said to yield consistent gains. Although those results belong to the later reinforcement stage, they reinforce the stated role of AC-DMD: it is the cold-start mechanism that makes the downstream reward-tilted optimization tractable under limited updates. Taken together, the empirical summary characterizes AC-DMD as a low-variance, scheduler-agnostic cold-start whose most visible standalone contribution is the consistency regularizer. The most direct evidence for that characterization is the reported improvement of over \(+1.5\) PickScore points before any RL is applied. Source: https://www.emergentmind.com/topics/ambient-consistent-distribution-matching-distillation-ac-dmd