Research map / Image generation / Autoregressive and latent generation
STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STARFlow, a scaled latent normalizing flow built on TARFlow with a deep-shallow Transformer design, latent-space learning with decoder fine-tuning, and a new classifier-free guidance for autoregressive flows, reaching near-diffusion quality on ImageNet and text-to-image generation.
Problem
Normalizing flows have received little attention for high-resolution, large-scale image synthesis; it was unclear whether TARFlow-style autoregressive flows could scale to competitive quality against diffusion and autoregressive models.
Approach
Scale TARFlow via a deep-shallow architecture (one deep Transformer block plus several shallow 2-layer blocks), learn the flow in the latent space of a pretrained autoencoder with noise injection and decoder fine-tuning, and use a principled Gaussian-derived guidance formula with variance-ratio clipping for classifier-free guidance.
Key results
- On class-conditional ImageNet 256x256, STARFlow (1.4B) reaches FID 2.40, versus 5.56 for TARFlow (1.3B) under the same codebase.
- Text-to-image zero-shot on COCO reaches FID 9.1 for the full-data 3.8B STARFlow, and 10.3 for the CC12M-trained 3.8B model.
- On GenEval, STARFlow scores 0.56 overall, slightly above SDXL (0.55) and above Emu3 (0.54).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace equal-sized flow blocks with a deep-shallow architecture in pixel space (ablation) | TARFlow · ImageNet 256×256 | FID-50K: 5.56 -> 4.69 | Helped |
| Move from pixel inputs to autoencoder latents (ablation) | Deep-shallow TARFlow · ImageNet 256×256 | FID-50K: 4.69 -> 2.40 | Helped |
| Finetune the decoder to reconstruct directly from noisy latents (ablation) | STARFlow · ImageNet 256×256 | FID-50K: 2.96 -> 2.40 | Helped |
| Use multi-step DiT denoising instead of single-step score denoising (ablation) | STARFlow · ImageNet 256×256 | FID-50K: 2.96 -> 2.53 | Helped |
| Increase training data from CC12M to the scaled in-house dataset (ablation) | STARFlow text-to-image · COCO 2017 val | FID-30K: 10.3 -> 9.1 | Helped |
| Initialize the deep block from Gemma2 and omit the additional text encoder (ablation) | STARFlow text-to-image · COCO 2017 val | FID-30K: 10.3 -> 11.4 | Hurt |
| Use a deep-shallow architecture instead of equal-sized flow blocks (ablation) | STARFlow · ImageNet 256×256 | inference time: 72 -> 35 | Helped |
| Increase the depth of the deep block (ablation) | STARFlow · ImageNet generation | FID | Helped |
| Use fewer than two layers in the deep block (ablation) | STARFlow · ImageNet generation | performance | Hurt |
| Increase the number of deep blocks rather than their depth (ablation) | STARFlow · ImageNet generation | performance | No clear effect |
| Replace the original autoregressive-flow CFG with the proposed score-based CFG (ablation) | STARFlow text-to-image · Text-to-image generation | FID | Helped |
Surprising findings
- Guidance is effective only in the top few flow blocks, with most effective compute concentrated in the block closest to the noise.
- Decoder fine-tuning on noisy latents outperforms single-step score denoising (FID 2.40 vs 2.96) and even multi-step DiT denoising (2.53).
Limitations
- Relies on frozen pretrained autoencoders, leaving joint latent-flow training unexplored.
- Inference speed is un-optimized and sequential generation is slow.
- Evaluation limited to class- and text-conditional image generation; no video or 3D.
Experimental setup
Datasets: ImageNet-1K CC12M in-house text-image dataset (~700M pairs) AnyEdit
Benchmarks: ImageNet 256x256 FID-50K ImageNet 512x512 FID-50K MSCOCO 2017 validation FID-30K GenEval
Metrics: FID NLL GenEval overall score inference throughput
Baselines: ADM LDM DiT SiT LlamaGen VAR MAR TARFlow
Builds on: TARFlow classifier-free guidance Stable Diffusion autoencoder MAR
Closest papers by meaning
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.91)
- STARFlow-V: End-to-End Video Generative Modeling with Autoregressive Normalizing Flows (CVPR 2026 · similarity 0.91)
- STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation (NeurIPS 2026 · similarity 0.89)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.89)
- Boosting Flow-based Generative Super-Resolution Models via Learned Prior (CVPR 2024 · similarity 0.88)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · 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.