Research map / Probabilistic and scientific modeling / Variational and generative sampling
SimFlow: Simplified and End-to-End Training of Latent Normalizing Flows
Qinyu Zhao, Guangting Zheng, Tao Yang, Rui Zhu, Xingjian Leng, Stephen Gould, Liang Zheng
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 |
|---|---|---|---|
| make VAE encoder trainable (end-to-end) with fixed variance instead of frozen VAE (ablation) | SimFlow · Variant study (80 epochs, gFID without CFG) | gFID: 67.41 -> 16.97 | Improved |
| train end-to-end with noise augmentation instead of frozen encoder (ablation) | NF with noise augmentation · Variant study (80 epochs, gFID without CFG) | gFID: 39.33 -> 20.01 | Improved |
| fix variance instead of noise augmentation under end-to-end training (ablation) | SimFlow · Variant study (80 epochs, gFID without CFG) | gFID: 20.01 -> 16.97 | Improved |
| fix VAE variance instead of learnable variance (frozen VAE) (ablation) | SimFlow · Variant study (80 epochs, gFID without CFG) | gFID: 84.44 -> 67.41 | Improved |
| fix VAE variance instead of learnable variance (frozen VAE) (ablation) | SimFlow · Variant study (80 epochs, gFID without CFG) | rFID: 0.23 -> 0.40 | Hurt |
| add REPA-E alignment to SimFlow | SimFlow · ImageNet 256x256 (class-conditional) | gFID (with guidance): 2.15 -> 1.91 | Improved |
| raise latent dimension from 64 to 128 (ablation) | SimFlow · Latent dimension sweep (gFID with CFG) | gFID: 2.53 -> 3.62 | Hurt |
| lower latent dimension from 64 to 16 (ablation) | SimFlow · Latent dimension sweep (gFID with CFG) | gFID: 2.53 -> 3.55 | Hurt |
| change fixed variance from 0.5^2 to 0.1^2 (ablation) | SimFlow · Fixed variance sweep (gFID with CFG) | gFID: 2.53 -> 3.01 | Hurt |
| change fixed variance from 0.5^2 to 0.75^2 (ablation) | SimFlow · Fixed variance sweep (gFID with CFG) | gFID: 2.53 -> 2.39 | No clear effect |
| scale NF from S to XXL (ablation) | SimFlow · NF model size sweep (gFID with CFG) | gFID: 15.34 -> 2.53 | Improved |
| use additive noise instead of linear noise (ablation) | SimFlow · Latent noise type (100K iterations, gFID without CFG) | gFID: 71.0 -> 39.6 | Improved |
Closest papers by meaning
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.90)
- Normalizing Flows with Iterative Denoising (ICML 2026 · similarity 0.90)
- Bidirectional Normalizing Flow: From Data to Noise and Back (CVPR 2026 · similarity 0.89)
- Denoising Normalizing Flow (NeurIPS 2021 · similarity 0.89)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.88)
- Normalizing Trajectory Models (NeurIPS 2026 · similarity 0.87)
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.
