Eyring, Karthik, Dosovitskiy, Ruiz and Akata (2025) — ARXIV-2508.09968
Key takeaways
- The problem is a latency figure. Reward-guided test-time noise optimization works — ReNO takes SANA-Sprint from 0.70 to 0.81 on GenEval — and costs 30 seconds a sample against the base model’s 0.2. Gradient-based variants backpropagate through the full model; gradient-free ones need thousands of denoiser evaluations. The paper’s framing is that this is test-time scaling’s general problem, met here in a tractable case.
- The move is to amortize it into weights, once. Train a lightweight network to predict the optimized initial noise directly, so inference is one extra forward pass rather than an optimization loop.
- Where the adaptation goes is the actual contribution. The target is a
reward-tilted distribution, and learning it needs a KL term to the base
model or the result reward-hacks. In data space that KL requires
Jacobian determinants through the generator — intractable. Moved into
noise space it becomes, under a Lipschitz condition, an
L2penalty on the magnitude of the noise modification, and the data-processing inequality makes it an upper bound on the data-space divergence you actually wanted. - The negative result is the sharp one. Direct LoRA reward fine-tuning of the same generator does not merely underperform — it degrades the base model, and degrades further with more sampling steps: SANA-Sprint GenEval 0.70 → 0.67 at one step, 0.72 → 0.66 at two, 0.73 → 0.62 at four. HyperNoise on the same model goes 0.70 → 0.75, 0.72 → 0.76, 0.73 → 0.77.
- What it recovers, stated honestly. SD-Turbo 0.49 → 0.57 (ReNO: 0.63). SANA-Sprint 0.70 → 0.75 (ReNO: 0.81, best-of-N: 0.79, LLM prompt optimization: 0.75). FLUX-schnell 0.68 → 0.72 (ReNO: 0.76). The authors’ own summary — “about half of the performance gains achieved by ReNO” — matches the numbers.
- Cost accounting is scrupulous. Added latency is 0.1–0.2 s; the tables count the hypernetwork pass as an extra NFE (a “one-step” model becomes NFEs = 2). 33× to 300× faster than the test-time methods it substitutes for.
- It is a few-step technique and the paper shows where it stops. The gain shrinks as NFEs grow — at 8 steps 0.74 → 0.76, at 16 steps 0.73 → 0.75, at 32 steps 0.71 → 0.72. Worth noting the base distilled model peaks around 4 steps and declines after, so both curves are flat where it matters.
Standing in the anthology
The other half of a trade the record only had one side of. SOTA-301,
filed from LIT-490, is the inference-time route: steer the sampler with
an objective gradient, correctly ordered, paying per sample. This is the
amortized route: pay once in post-training, pay almost nothing per sample,
and recover roughly half. compared_against: LIT-490 is declared because the
two answer the same question with opposite cost profiles and neither
dominates.
It also lands beside SOTA-184 with an unusual twist: LoRA is the mechanism in both, and here it adapts the input rather than the model, which is what makes the regularizer tractable.