Cannot reproduce "without CFG" FID in Table 3 (expected ~4.64, got ~12), while "with CFG" is reproducible

[Lil-Shake]

Hi, thanks for the great work.

I’m trying to reproduce the ablation results and have a mismatch specifically on the without-CFG setting.

What I can reproduce

With CFG settings, my results are close to the reported trend/quality (so the overall pipeline seems correct).

What I cannot reproduce

For without CFG (CFG disabled), Table 3 reports around FID = 4.64 (VAR-d16, Top-k shown as 600 in the table), but I get about FID = 12.11.

My setup

I followed the sampling/eval workflow according to mar/README.md (same general procedure for sampling + evaluator).

• Reference batch: VIRTUAL_imagenet256_labeled.npz
• Evaluator: OpenAI guided-diffusion/evaluator.py

Observed Results

Inception Score: 106.05
FID:       12.11
sFID:      9.42
Precision: 0.709
Recall:    0.588

(Command with CFG gives expected behavior close to paper trend, omitted here for brevity.)

Questions

Could you help clarify what exact config is required for the without-CFG row in Table 3? In particular:

1. Should Top-k be exactly 600 and Top-p be disabled?
2. Any required seed/sampler/distributed settings for this ablation?
3. Any preprocessing/evaluator detail that differs from README defaults?
4. Is the “without CFG” row generated from the same checkpoint and code path as the public release?

Thanks a lot!