Martin Garcia, Knaebel, Schmidt, de Geus, Hermans and Leibe (2024; WACV 2025) — ARXIV-2409.11355
first_author follows arXiv’s metadata (“Garcia, Gonzalo Martin”). The
surname may be the double “Martin Garcia”.
Key takeaways
The bug. “for a single-step prediction, the model receives a timestep
encoding that indicates an almost perfect depth map whereas the actual input
is pure noise.” DDIM’s leading spacing “excludes the final timestep T”.
With T = 1000, one leading step is [1], and one trailing step is [1000].
“In the limit of k → T inference steps, both strategies converge.” So the
bug bites exactly where few-step inference lives.
The fix is borrowed, and the paper says so. “we can use the trailing setting as proposed in recent work [28]”, which is Lin et al. 2023 (arXiv 2305.08891). “this bug has already been reported in the general diffusion model literature”. It takes only the trailing spacing from Lin et al., and not the zero-terminal-SNR rescaling. SD’s schedule is untouched. The contribution is showing that the fix, “slight” for image generation, is “crucial for single-step inference in models such as Marigold”.
End-to-end fine-tuning. Fix t = T. Feed zeros as the noise (“the mean of
the noise distribution”). Convert the v-prediction to ẑ₀, decode through the
frozen VAE, and train on the task loss: affine-invariant L1 for depth,
angular error for normals. 20K iterations on the same 74K synthetic samples
Marigold used, about 3 days on one H100. Fine-tuning the decoder too “does
not improve performance”.
Depth, zero-shot (AbsRel ↓):
| method | steps × ens. | NYUv2 | KITTI | ETH3D | ScanNet | DIODE |
|---|---|---|---|---|---|---|
| Marigold | 50 × 10 | 5.5 | 9.9 | 6.5 | 6.4 | 30.8 |
| Marigold | 50 × 1 | 6.0 | 10.5 | 7.1 | 6.9 | 31.0 |
| Marigold + trailing fix | 1 × 1 | 5.7 | 10.8 | 6.9 | 6.6 | 31.1 |
| Marigold + E2E FT | 1 × 1 | 5.2 | 9.6 | 6.2 | 5.8 | 30.2 |
| SD + E2E FT | 1 × 1 | 5.4 | 9.6 | 6.4 | 5.8 | 30.3 |
Normals (mean angular error ↓): Marigold (50 × 10) 18.8 / 17.7 / 18.4 / 39.1 on NYUv2 / ScanNet / iBims-1 / Sintel. The fixed single step gives 17.4 / 16.8 / 18.1 / 37.1, and E2E fine-tuning 16.2 / 14.7 / 15.8 / 33.5.
After the fix, more denoising steps make it worse. “instead of improving the depth map with more denoising steps, the performance actually gets worse”. The offered reason is that the model “expects noised ground truth latents instead of its own predictions”. That reason is not tested, and the curve is shown only as a figure.
Traps
- “Single-step matches multi-step” is true only against the un-ensembled run. Against Marigold’s 50-step, 10-member ensemble, the fixed single step is worse on AbsRel on all five depth datasets. Against 50 steps without ensembling it wins three of five. The abstract’s “performs comparably” needs that qualifier. What beats the ensemble everywhere is E2E fine-tuning, a training change, and not the scheduler fix.
- Not zero terminal SNR. This paper is evidence for Lin et al.’s spacing fix and says nothing about its schedule rescaling. SOTA-263’s zero-terminal-SNR discussion should not cite it.
- “>200× faster” compares 500 network evaluations (50 steps × 10) with one.
- On normals, one metric disagrees. The fixed single step wins on mean angular error everywhere, but on the 11.25° accuracy metric it loses on ScanNet and iBims-1.
- “SD is as good as Marigold’s init” rests on 0.0–0.2 AbsRel at one fine-tuning budget, with no seeds. The normals baseline is a Marigold model the authors trained themselves.
Standing in the anthology
Filed from the reading-time triage of 2026-09-25. That entry summarized this paper as “once that is fixed single-step inference matches it”. That is the trap above, and the summary here replaces it.
It sources two practices:
SOTA-416: use trailing timestep spacing for few-step sampling. This is a single-variable measurement: same weights, same data, only the scheduler changed.SOTA-414: to make an image diffusion model a dense geometric predictor, fine-tune it end to end as a one-step model with a task loss.
The two papers it depends on are now held. It corrects Marigold (LIT-691) and extends Lin et al. (LIT-689, WACV 2024; this paper’s bibliography says “WACV, 2023”). Two of this note’s readings of Lin were corrected on filing it. Lin credits the trailing discretization to DPM-Solver rather than proposing it. And “slight” is this paper’s characterization, which Lin, who never measures spacing alone, neither states nor supports. It has no direct relation to LIT-107 (MiDaS v3.1). The paper compares against the original MiDaS, not v3.1.