Source
Karras, Aittala, Kynkäänniemi, Lehtinen, Aila and Laine (2024), LIT-721 — ARXIV-2406.02507. THEORY-112 is why it works.
What to do
Train, or keep, a deliberately worse version of the model you are going to sample. Same task, same conditioning, same data distribution — degraded on capacity and training length. The paper’s working configuration for EDM2-S: an XS-sized guiding model at 1/16 of the main model’s training iterations. Both degradations together beat either alone.
Guide with it in place of the unconditional model, using the ordinary guidance formula and a weight found by search.
Give the two models independent EMA lengths. Forcing them equal costs FID 1.34 → 1.53. That makes SOTA-428’s post-hoc EMA a precondition for the full benefit rather than an optional companion.
The condition that actually binds
The two models must be degraded in the same way. This is not a caveat, it is the mechanism, and the paper tested it by breaking it:
D₁, D₀ from one base model (FID 2.56) | autoguided |
|---|---|
| dropout 5% (4.98) and 10% (15.00) | 2.55 at w = 2.25 |
| input noise +10% (3.96) and +20% (9.73) | 2.56 at w = 2.00 |
| dropout on one, input noise on the other | nothing works; best w is 1 |
A worse reference model is not sufficient. It has to be worse in the same direction, so that the difference between the two predictions points along the error the good model is making. Capacity and training length are recommended precisely because any model already suffers from both to some degree.
Do not read the synthetic corruptions as a recipe. The authors are explicit that they do not suggest guiding with dropout or input noise in practice; those were constructed to test the hypothesis, and a real model does not have those particular defects.
What it buys
The recommendation originates in LIT-721, which proposed autoguidance (guiding the conditional model with a smaller, less-trained copy of itself where CFG uses an unconditional model) and argued that CFG’s quality gain comes from the reference model being worse, not from class emphasis. All of its evidence is on EDM2 checkpoints and measured by Fréchet distances, which is where every number below comes from:
| setting | baseline | autoguided |
|---|---|---|
| ImageNet-512, EDM2-S | 2.56 | 1.34 |
| ImageNet-512, EDM2-XXL | — | 1.25 |
| ImageNet-64 | — | 1.01 |
| ImageNet-512, unconditional EDM2-S | 11.67 | 3.86 |
FD_DINOv2, ImageNet-512 | 29.16 | 24.18 |
The 1.34 also beats the concurrent CFG + Guidance Interval’s 1.68, and the two were not found to combine.
The unconditional row is the one with no CFG analogue. SOTA-424 requires a condition to drop; this does not, so it reaches a regime that practice explicitly excludes.
Costs and conditions
You need a second network at sampling time, as with CFG — but a small one, which is cheaper than CFG’s full-size unconditional model. Against that, you need to have kept an early small snapshot, and the authors name this as the practical blocker: such snapshots are “not available for current large-scale image generators in practice”, and staged training whose data changes partway “would violate our assumptions”. If your pipeline does not already retain them, the cost is a training run, not a checkpoint.
The diversity claim is measured with mixed metrics. FID and FD_DINOv2 both
combine fidelity and coverage, so “quality without losing variation” currently
rests on a metric that cannot separate them — exactly what SOTA-425 exists to
say. The paper’s own future work asks for precision and recall here and does not
report them. Treat preserved diversity as the hypothesis, not the finding.
Guidance-weight and EMA search is part of the method, not tuning noise. EDM2 networks are “known to be sensitive to the guidance weight and EMA length”, and the results come from a grid search over both. SOTA-307 and SOTA-425 apply to reading any of these numbers.
High-noise schedules may be unnecessary here. Noise-level-dependent guidance weights exist largely to suppress CFG where differently-conditioned distributions diverge; the authors expect no such problem “as both models target the same distribution”. Untested.
Related
- SOTA-424 — CFG. This replaces its reference model and thereby its trade; two of that practice’s Conditions are CFG-specific rather than general, and it now says so.
- SOTA-428 — post-hoc EMA. A precondition for the best results here.
- SOTA-425 — report precision and recall beside FID. The measurement this practice’s central claim is waiting on.
- SOTA-397 — conditioning an unconditionally trained model. Different problem: that adds a condition, this improves an unconditional model without one.