Research map / Image generation / Text-to-image diffusion
EchoGen: Generating Visual Echoes in Any Scene via Feed-Forward Subject-Driven Auto-Regressive Model
Ruixiao Dong, Zhendong Wang, Keli Liu, Li Li, Ying Chen, Kai Li, Daowen Li, Houqiang Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DINOv2 instead of SigLIP-2 for fine-grained semantic injection (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.438 -> 0.632 | Improved |
| use DINOv2 instead of FLUX VAE features for semantic cross-attention (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.433 -> 0.632 | Improved |
| add global semantic prefix token with AdaLN (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.632 -> 0.670 | Improved |
| add global semantic prefix token with AdaLN (ablation) | EchoGen-0.1B · DreamBench | CLIP-T: 0.328 -> 0.322 | Mixed |
| inject semantic features with cross-attention instead of multi-modal attention (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.646 -> 0.670 | Improved |
| add content features via multi-modal attention (ablation) | EchoGen-0.1B · DreamBench | CLIP-I: 0.798 -> 0.806 | Improved |
| add content features via cross-attention instead of MM-Attn (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.670 -> 0.667 | No clear effect |
| add subject segmentation (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.663 -> 0.672 | Improved |
| add subject segmentation at inference (ablation) | EchoGen-2B · DreamBench | DINO: 0.737 -> 0.755 | Improved |
| prepend global semantic feature instead of content feature (ablation) | EchoGen-0.1B · DreamBench | DINO: 0.663 -> 0.672 | Improved |
| increase subject guidance weight in CFG (ablation) | EchoGen · DreamBench | CLIP-I vs CLIP-T | Mixed |
| use VAR backbone instead of diffusion for feed-forward subject generation | EchoGen-2B · DreamBench | sampling latency: 5.2s vs 25.4s (EasyControl) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- EasyControl — EasyControl: Adding Efficient and Flexible Control for Diffusion Transformer (ICCV 2025) · better on 4 of 5 shared results
- OminiControl — OminiControl: Minimal and Universal Control for Diffusion Transformer (ICCV 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CART: Compositional AutoRegressive Transformer for Image Generation (AAAI 2026 · similarity 0.88)
- EchoShot: Multi-Shot Portrait Video Generation (NeurIPS 2025 · similarity 0.88)
- DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation (CVPR 2023 · similarity 0.87)
- VersaGen: Unleashing Versatile Visual Control for Text-to-Image Synthesis (AAAI 2025 · similarity 0.87)
- OneActor: Consistent Subject Generation via Cluster-Conditioned Guidance (NeurIPS 2024 · similarity 0.86)
- DORSal: Diffusion for Object-centric Representations of Scenes $\textit{et al.}$ (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
