FastGRN cuts GRN image generation from 50 refinement steps to 2, matching its teacher's ImageNet FID at 12.7 times the speed
Synopsis
Discrete-token image generator GRN needs 50 refinement steps per image, and cutting the step count makes its samples fall apart. FastGRN treats GRN's own uniform-state corruption as a discrete diffusion kernel, computes a per-token reverse-KL gradient directly from two posterior networks, and adds windowed matching, oversampling and non-uniform sampling so that a 2-step student matches the FID of its 79-network-call teacher on ImageNet while running 12.7 times faster per image.
Interpretation
The sampling cost of the discrete-token generator GRN drops from 50 refinement steps to 2 without the image quality collapsing. GRN reaches high fidelity with 50 refinement steps plus classifier-free guidance, 79 network calls per image; simply taking 4 of those 50 steps makes ImageNet samples fall apart, and later steps cannot repair them. On ImageNet, the 2-step FastGRN-B reaches FID-50k 3.73, level with its teacher, with network calls cut from 79 to 2; the 3-step version reaches 3.42, below the teacher.
Few-step failure comes from a joint error between tokens rather than from inaccurate single-token prediction. GRN's output head is factorized, so per-token cross-entropy training can reach at best the product of posterior marginals, and the remaining error is the conditional total correlation of the tokens given the current state. Once GRN's uniform-state corruption is read as a discrete diffusion kernel, the per-token reverse-KL gradient follows directly from the concrete scores of a real and a fake posterior network, with no policy gradient, straight-through estimator or discriminator.
Training has three parts: a window with random steps, oversampling with non-uniform sampling, and refinement confidence keeping. Without the window, 2-step matching starts healthy and then collapses into blank images; a fixed-step student specializes to one budget and drifts when the step count changes. The window restricts matching to moderate noise levels and keeps training stable; random steps let one student cover 4 to 8 steps at flat quality; oversampling fine-tunes the judge on the student's own extra steps, and non-uniform sampling raises the second step's share of the 2-step path to 75%.
Perspective
The demonstrated scope is class-conditional ImageNet generation with GRN-B (130M) and GRN-L (458M), plus text-to-image on the public GRN-T2I 2B checkpoint; the text-to-image students use windowed matching alone, without the other parts. The authors propose applying the same parts to bit-level text-to-image and video models as the next step.
On text-to-image the 2-step student keeps 97% of the teacher's GenEval, and the 4-step student scores 0.689 against the teacher's 0.695, so it does not fully close the gap; on GRN-B non-uniform sampling moves 2-step FID-50k from 3.83 to 3.73, a smaller gain than on GRN-L. Confidence keeping is weakly predictive on a 2-step draft, with AUC 0.54, and helps only when used during training.
