Research map / Image generation / Autoregressive and latent generation
FlowTok: Flowing Seamlessly Across Text and Image Tokens
Ju He, Qihang Yu, Qihao Liu, Liang-Chieh Chen
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 |
|---|---|---|---|
| learn the text alignment target with an MLP instead of average pooling (ablation) | FlowTok-B · COCO FID-30K | FID-30K: 36.02 -> 29.14 | Improved |
| use contrastive text alignment loss instead of cosine similarity loss (ablation) | FlowTok-B · COCO FID-30K | FID-30K: 31.80 -> 29.14 | Improved |
| raise text alignment loss weight from 1.0 to 2.0 (ablation) | FlowTok-B · COCO FID-30K | FID-30K: 29.14 -> 30.59 | Hurt |
| scale FlowTok-XL (698M) to FlowTok-H (1.1B) | FlowTok · MJHQ-30K | FID: 7.68 -> 7.15 | Improved |
| scale FlowTok-XL (698M) to FlowTok-H (1.1B) | FlowTok · COCO FID-30K | FID-30K: 10.06 -> 9.67 | Improved |
| use compact 1D text-to-image flow instead of 2D CrossFlow | FlowTok-H · text-to-image throughput | throughput (samples/s): 1.1 -> 18.2 | Improved |
| use compact 1D tokens instead of 2D CrossFlow | FlowTok-H · training cost | 8 A100 days: 78.8 -> 26.1 | Improved |
| use 1D flow instead of CrossFlow on COCO | FlowTok-H · COCO FID-30K | FID-30K: 9.63 -> 9.67 | No clear effect |
| use direct image-to-text 1D flow instead of CrossFlow | FlowTok-XL · COCO Karpathy | CIDEr: 116.2 -> 117.0 | Improved |
| use direct image-to-text flow instead of SCD-Net | FlowTok-XL · COCO Karpathy | CIDEr: 118.0 -> 117.0 | Hurt |
| use 1D flow over text-as-condition Stable-Diffusion-2.1 | FlowTok-H · MJHQ-30K | FID: 26.96 -> 7.15 | 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.
- GLIDE — GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models (ICML 2022) · better on 1 of 1 shared result
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
- Flowing from Words to Pixels: A Noise-Free Framework for Cross-Modality Evolution (CVPR 2025 · similarity 0.89)
- FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching (ICML 2025 · similarity 0.88)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.87)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.87)
- Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
- End-to-End Training for Unified Tokenization and Latent Denoising (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.
