Research map / Image generation / Autoregressive and latent generation
Self-transcendence: Is External Feature Guidance Indispensable for Accelerating Diffusion Transformer Training?
Lingchen Sun, Rongyuan Wu, zhengqiang ZHANG, Ruibin LI, Yujing Sun, Shuaizheng LIU, Lei Zhang
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 |
|---|---|---|---|
| add VAE structure guidance (ablation) | SiT-B/2 · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 38.60 -> 32.20 | Improved |
| add self-guided representation (ablation) | SiT-B/2 · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 38.60 -> 25.21 | Improved |
| add self-guided representation on top of VAE structure guidance (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 32.20 -> 22.91 | Improved |
| add VAE structure guidance to the guiding model (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 25.21 -> 22.91 | Improved |
| use two-stage instead of one-stage training (ablation) | Self-Transcendence · ImageNet 256 no CFG (SiT-B/2) | FID: 34.50 -> 20.49 | Improved |
| use VAE-aligned guiding model instead of REPA-trained teacher (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 23.40 -> 22.91 | Improved |
| raise guidance scale from 1.0 to 30.0 (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 29.30 -> 22.91 | Improved |
| raise guidance scale from 30.0 to 60.0 (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 22.91 -> 23.57 | Hurt |
| train guiding model 300K instead of 200K iterations (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 22.91 -> 23.05 | No clear effect |
| train guiding model 200K instead of 50K iterations (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 28.32 -> 22.91 | Improved |
| guide layer 6 from layer 8 instead of the same layer 6 (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 24.37 -> 22.91 | Improved |
| guidance weight 0.5 instead of 0.1 (ablation) | Self-Transcendence · ImageNet 256 ablation (SiT-B/2, 80 ep) | FID: 24.28 -> 22.91 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- + LayerSync — LayerSync: Self-aligning Intermediate Layers (ICLR 2026) · better on 2 of 2 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
- SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.90)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.90)
- DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing (NeurIPS 2026 · similarity 0.88)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.88)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.88)
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · similarity 0.88)
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.
