* Both authors contributed equally to this research.
Image analogy copies a change shown by a before/after pair (A, A′) onto a new image B. But one pair usually holds several edits at once, and existing methods transfer all of them — you cannot ask for only some.
We introduce Selective Image Analogy (SIA): the pair stays fixed, and you choose which of its edits to apply. SIA first learns to transfer the whole change, then gains control from a mask over the edit words. The mask acts twice — it weakens the unwanted words in the text embedding, and gates a small adapter to them while the image is generated. The unwanted edits are dropped and the chosen ones stay, even though every edit is still visible in A′, and no new pair is needed for each combination.
Every edit starts kept — click one to suppress it. The demonstration pair never changes, only the result. Every image here is a real SIA output.
One suppression setting per example. Drag the slider to wipe between the untouched query image and the result.
Before denoising the mask weakens the suppressed tokens in the text embedding; during denoising it gates the adapter to those same tokens. The demonstration pair and the prompt stay untouched.
A lightweight LoRA adapter on FLUX.2-Klein, trained in concat mode on the four-image tuple (A, A′, B, B′). The latents of A, A′ and B are packed as visual conditioning and paired with a structured analogy prompt that names all demonstrated edits. After this stage the model reliably performs full transfer.
A token-gated LoRA whose low-rank residual is applied only at the text positions marked by the suppression mask. Unmasked tokens pass through the ordinary text stream, so suppression stays local to the edit phrase the user switched off — no new demonstration pair, no per-subset prompt, no query-specific optimisation.
The prompt embeddings still describe every edit, suppressed ones included. For each edit we estimate a direction in text-encoder space from 64 contrastive positive/negative prompt pairs, cache it, and subtract that component from the suppressed edit’s token embeddings before denoising. Retained edits and the analogy instructions are left untouched.
Our Multi-Edit Analogy Dataset spans creature and non-creature subjects and covers object additions, appearance changes, expression changes, pose changes, background replacements and style transfer. Every sample is filtered with a vision–language model to verify that the desired edits are present and the excluded ones are absent.
Preservation alone is not enough for selective analogy: a method that barely changes the query image scores well on preservation and badly on everything else. SIA leads on retained accuracy, recall and both CLIP directional scores.
| Method | Qwen3-VL-32B | Gemma 3 | CLIP | |||||
|---|---|---|---|---|---|---|---|---|
| Ret. Acc. ↑ | Pres. ↑ | Recall ↑ | Ret. Acc. ↑ | Pres. ↑ | Recall ↑ | Dir. Image ↑ | Dir. Ret. ↑ | |
| RelationAdapter [Gong et al. 2025] | 6.74 | 9.93 | 0.782 | 8.09 | 9.48 | 0.941 | 0.386 | 0.144 |
| VisualCloze [Li et al. 2025] | 7.24 | 8.62 | 0.856 | 8.14 | 9.28 | 0.947 | 0.464 | 0.152 |
| Edit Transfer [Chen et al. 2025] | 2.31 | 9.96 | 0.176 | 5.75 | 9.58 | 0.623 | 0.102 | 0.023 |
| LoRWeB [Manor et al. 2026] | 6.59 | 9.83 | 0.736 | 8.12 | 9.51 | 0.926 | 0.402 | 0.155 |
| SIA (ours) | 8.01 | 9.75 | 0.892 | 8.65 | 9.57 | 0.952 | 0.543 | 0.171 |
@inproceedings{alimohammadi2026sia,
title = {SIA: Selective Image Analogy},
author = {Alimohammadi, Amirhossein and Izadpanah, Kian and
Kariminia, Bardiya and Chrysanthou, Yiorgos and
Mahdavi-Amiri, Ali},
booktitle = {SIGGRAPH Asia 2026 Conference Papers},
year = {2026},
publisher = {Association for Computing Machinery}
}
Questions and requests are welcome — please open an issue on the GitHub repository.