FMRG Beats Top Image Guidance Methods 70x Faster With 3 Steps
FMRG reformulates generative model guidance as deterministic optimal control, achieving 70x speedups over prior art with as few as 3 network evaluations
Takeaways
- What: FMRG reformulates generative model guidance as deterministic optimal control, replacing the standard reward-tilting framework with a principled single-trajectory approach.
- Speed: Achieves state-of-the-art results with as few as 3 NFEs (network function evaluations), delivering up to a 70× speedup over prior best methods like FMTT.
- Unification: Proves that DPS, FlowDPS, FlowChef, MPGD, ReNO, and D-Flow are all special cases of FMRG — each approximating the same optimal control signal at different levels of coarseness.
- Two variants: FMRG-J (Jacobian, on-manifold, ~48 GB VRAM) for complex rewards; FMRG-E (Euclidean, single L40S GPU) for reconstruction tasks.
- Scope: Tested on inverse problems, style transfer, human preference rewards, and 7B VLM rewards on FLUX-scale text-to-image models.
- Code: Training-free, open-source at github.com/jrrhuang/fmrg; requires a pre-trained flow map (FLUX.1-Dev LoRA checkpoint provided).
Steering a generative model toward a specific reward , better aesthetics, sharper reconstructions, stronger prompt alignment , is one of the most practically important problems in modern image synthesis. A new paper by Jerry Huang, Justin Lin, Sheel Shah, Kartik Nair, and Nicholas Boffi proposes Flow Map Reward Guidance (FMRG), a framework that rethinks this problem from first principles and delivers results that are both theoretically cleaner and dramatically faster than existing approaches.
Why existing guidance is broken
In generative modeling, we often wish to produce samples that maximize a user-specified reward such as aesthetic quality or alignment with human preferences , a problem known as guidance. Despite their widespread use, existing guidance methods either require expensive multi-particle, many-step schemes or rely on poorly understood approximations. The dominant theoretical framing casts guidance as sampling from a reward-tilted distribution , essentially, the base model's distribution reweighted by how much reward each sample would earn. The math is clean, but the computation is brutal: exact sampling from this tilted distribution is intractable in high dimensions, and methods like Sequential Monte Carlo (SMC) that try to approximate it need many particles running in parallel across many steps. Single-trajectory shortcuts like Diffusion Posterior Sampling (DPS) sidestep the ensemble but rely on approximations whose statistical properties are poorly characterized.
A second, underappreciated problem is that the reward-tilt framework is rooted in stochastic processes , SDEs, score functions, Doob h-transforms , while modern systems like FLUX and Stable Diffusion 3 increasingly use deterministic probability flows. There is no principled guidance methodology native to the deterministic, few-step, single-trajectory regime that these models actually use. FMRG is the first such framework.
Optimal control instead of reward tilting
The authors reformulate guidance as a deterministic optimal control problem, yielding a hierarchy of algorithms that subsumes existing approaches at the coarsest level. They show that the flow map , an object of significant recent interest for its role in fast inference , arises naturally in the optimal solution.