Fine-Tuning-Free Reference-Guided Diffusion for Structure-Preserving Elongated Fabric Defect Generation

. 2026 Oct 8;26(19):6351. doi: 10.3390/s26196351

Algorithm 1 Fine-tuning-free inference process of SLAID

Require:
Frozen AnyDoor model $D_{θ}$ , DDIM sampler, reference image $I_{r}$ , reference mask $M_{r}$ , target image $I_{t}$ , target mask $M_{t}$
Ensure:
Synthetic elongated fabric defect image $\hat{I}$

1:
Estimate defect geometry from $M_{r}$ .
2:
Build shape-aware reference representations $R$ using SSRM.
3:
Extract adaptive high-frequency guidance H using AHFG.
4:
Compose target control hint $C_{t}$ from $I_{t}$ , $M_{t}$ , and H.
5:
Encode reference representations into conditions $c_{r}$ .
6:
Encode target crop into latent $z_{t}^{0}$ and construct anchor mask $M_{anc}$ .
7:
Initialize DDIM latent $x_{T}$ with target-domain anchoring.
8:
for $τ = T$ to 1 do
9:
Predict noise $ϵ_{θ} (x_{τ}, C_{t}, c_{r})$ with the frozen AnyDoor model.
10:
Apply shape-aware detail guidance from $R$ .
11:
Perform one DDIM denoising step to obtain $x_{τ - 1}$ .
12:
Anchor non-edited regions in latent space using TLA.
13:
end for
14:
Decode $x_{0}$ and blend the generated patch back into $I_{t}$ .
15:
return $\hat{I}$ .