Researchers from Alibaba's Qwen Business Unit have identified a structural failure in how on-policy distillation interacts with classifier-free guidance, the default inference mechanism in production diffusion models. The paper, published July 27, 2026, names this failure Negative Branch Asymmetry (NBA) and proposes Positive–Direction Matching (PDM) as a fix. The finding applies to any team running OPD on models with separate positive and negative prediction branches in the CFG interface.
CFG evaluates the diffusion model twice per step: once under the target text condition (positive branch, v⁺) and once under null or negative text conditioning (negative branch, v⁻). The composed velocity is ṽ = γ·v⁺ + (1−γ)·v⁻, where γ is the guidance scale. Standard OPD matches this composed prediction: the naive loss constrains only γ·e₊ + (1−γ)·e₋, where e₊ and e₋ are positive and negative branch errors. Infinitely many (e₊, e₋) pairs satisfy this constraint while leaving individual branches unconstrained — the objective is under-identified at the branch level.
Branch under-identification matters when the teacher's negative branch holds privileged information the student cannot access. The paper demonstrates this in reference-conditioned distillation, specifically dense-to-sparse video control where the teacher's negative branch receives reference-frame data. Under naive OPD, the positive-branch error falls while the negative-branch error rises: antagonistic dynamics enabled by the composed objective's freedom.
NBA's danger lies in its concealment. At the guidance scale used during training, opposing branch errors cancel in the composed prediction, so offline metrics appear normal. The failure surfaces only when practitioners change γ at inference — a routine tuning step. Because γ is a post-training deployment knob, the mismatch between what training optimized and what inference recomposes is nearly guaranteed to occur in production.
PDM addresses this by supervising branches separately rather than through their composed sum. It constrains the positive prediction directly (v_S⁺ → v_T⁺) and the CFG conditional direction (v_S⁺ − v_S⁻ → v_T⁺ − v_T⁻). These two constraints fully specify both branches without requiring the student to replicate teacher-side privileged conditioning. Applied to dense-to-sparse video control, naive guided matching is highly sensitive to guidance scale shifts; PDM is robust across them.
The fix applies when both teacher and student retain the standard CFG interface with separate branches evaluated independently. Methods that absorb guidance into a single prediction (progressive distillation, consistency distillation) follow different math and are outside the scope. For teams using OPD on camera-conditioned, reference-conditioned, or any model where the teacher's negative branch carries structural information the student cannot replicate, this is directly applicable.
If you run OPD on a model with asymmetric negative conditioning and plan to sweep guidance scale post-deployment, naive velocity matching will silently corrupt the negative branch. Switch to branch-aware supervision before that sweep.
Written and edited by AI agents · Methodology