Text-to-Image & Video Generation/Editing
Dense Motion Supervision Without Optical-Flow Ground Truth
Generated automatically from the limitations stated in 8 papers (CVPR, ICCV, ICML), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Nearly every pipeline that needs dense correspondence—video generation control, camera pose estimation, motion attribution—currently anchors its supervision or its inference to optical flow, either ground-truth annotations or outputs of a pretrained model (RAFT, AllTracker). This couples training data to static-camera, high-quality, curated footage, caps dataset size (22k–630k pairs), injects flow errors directly into downstream quality (cycle-consistency errors >1000 px), and bakes in simplifying assumptions like first-frame conditioning or pinhole intrinsics. Every consumer of flow inherits the same ceiling set by flow estimators and the data they were trained on.
Why it matters
Video generation, pose/SLAM, and motion-attribution systems trained on vastly larger, uncurated video without inheriting flow-estimator errors; flow stops being a single point of failure.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Build a correspondence benchmark with per-domain reliability labels: run existing flow estimators (RAFT, AllTracker, video-depth-induced correspondence) on diverse domains (egocentric, cartoon, camera-only motion, dynamic scenes) and measure agreement against multi-teacher consensus plus held-out hand annotations. Deliverable: a dataset and a per-pixel confidence map showing where flow supervision is trustworthy. Measured: consensus error vs. domain, correlation of confidence with downstream task error.
- 2
Train downstream models (e.g., video generation control or pose estimation) with confidence-weighted flow supervision, where unreliable flow pixels are down-weighted or replaced by self-supervised consistency losses (cycle consistency, photometric warping). Measured: downstream quality as a function of how much filtered/confident flow is retained, compared to hard-thresholded filtering pipelines.
- 3
Distill flow supervision from self-supervised objectives only—train a correspondence model using cycle-consistency over unlabeled video and evaluate whether it substitutes for RAFT/AllTracker outputs in the pipelines above. Measured: downstream task parity against flow-anchored baselines with no flow model in the loop.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Self-supervised or confidence-weighted correspondence may simply be too weak where optical flow is genuinely needed—precise, dense, large-displacement matching—and downstream tasks may end up needing full curated flow supervision anyway.
Sub-problems
Self-Supervised Training and Calibration for Flow-Driven Camera Pose Estimation
Current learning-based camera pose and Structure-from-Motion architectures rely strictly on optical flow backbones trained with synthetic ground-truth annotations or fixed off-the-shelf flow estimators. When deployed on in-the-wild capture domains such as egocentric video, these models fail because real-world optical characteristics, lens distortions, and dynamic artifacts diverge sharply from synthetic training distributions. Without self-supervised training formulations, practitioners cannot fine-tune flow-driven pose models directly on unlabelled target domain videos where ground-truth optical flow is physically impossible to obtain.
Evidence
Each paper's own statement of the limitation, verbatim.
- Motion Attribution for Video GenerationICML 2026
Requires accurate optical flow from AllTracker; attribution quality depends on motion-detection quality, and degenerate clips (camera-only motion, cartoons) are filtered by assumption rather than validated per domain
- Learning to Generate Highly Dynamic Videos using Synthetic Motion DataCVPR 2026
Relies on estimated optical flow for real video supervision, which contains errors (max cycle-consistency error 1080.05 pixels) partially mitigated only by filtering
- Tora: Trajectory-oriented Diffusion Transformer for Video GenerationCVPR 2025
Requires a curated pipeline (motion segmentation, camera-motion detection, optical-flow filtering) and manually annotated internal videos; only ~630k eligible videos selected
- MotiF: Making Text Count in Image Animation with Motion Focal LossCVPR 2025
Requires RAFT optical-flow computation on training videos and a proprietary 1M video-text internal training dataset; the condition image is assumed to be the first frame
- Learning Normal Flow Directly From EventsICCV 2025
Depends on ground-truth optical flow for supervised training; no self-supervised variant, which the authors note limits usability
- FramePainter: Endowing Interactive Image Editing with Video Diffusion PriorsICCV 2025
Training data curation requires static-camera, high-quality videos and filtering by optical-flow magnitude, restricting supervision to natural motion patterns and yielding only ~22k pairs
- DifFlow3D: Toward Robust Uncertainty-Aware Scene Flow Estimation with Iterative Diffusion-Based RefinementCVPR 2024
Requires ground truth flow annotations for training; evaluated fully-supervised in main results
- FlowCam: Training Generalizable 3D Radiance Fields without Camera Poses via Pixel-Aligned Scene FlowNeurIPS 2023
Depends on off-the-shelf optical flow quality and known or simplistic predicted intrinsics (single global field-of-view, centered principal point); does not model radial distortion (e.g., Ego4D)
Nearest existing work
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video SynthesisCVPR 2024
- Learning by Analogy: Reliable Supervision From Transformations for Unsupervised Optical Flow EstimationCVPR 2020
- Unpaired Visual Editing with Self-Consistent Flow MatchingICML 2026
- FlowMotion: Training-Free Flow Guidance for Video Motion TransferCVPR 2026
- Self-Supervised Learning via Conditional Motion PropagationCVPR 2019
- DF-Net: Unsupervised Joint Learning of Depth and Flow using Cross-Task ConsistencyECCV 2018
- Learning Optical Flow From Still ImagesCVPR 2021
- Learning Accurate Dense Correspondences and When To Trust ThemCVPR 2021
- MotionAgent: Fine-grained Controllable Video Generation via Motion Field AgentICCV 2025
- U^2Flow: Uncertainty-Aware Unsupervised Optical Flow EstimationCVPR 2026
- AutoFlow: Learning a Better Training Set for Optical FlowCVPR 2021
- FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video TranslationCVPR 2024
- RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion TransformersNeurIPS 2025
- MotionV2V: Editing Motion in a VideoCVPR 2026
- AnimateAnything: Consistent and Controllable Animation for Video GenerationCVPR 2025
Related open problems
Text-to-Image & Video Generation/Editing
Systematically Benchmarking and Calibrating VLM-as-a-Judge Failure Modes in Image and Video Generation
Current text-to-image and video generation benchmarks overwhelmingly rely on Vision-Language Models (VLMs) like GPT-4o, Gemini, and Qwen-VL as automated judges for win rates and task success metrics. However, these VLM judges systematically exhibit low correlation with human raters on fine-grained visual details, suffer from position and same-family evaluator biases, and under-penalize visual artifacts such as boundary glitches or text corruption. Consequently, benchmark progress across generative models is structurally uninterpretable, as claimed state-of-the-art gains may reflect evaluator artifacts rather than genuine improvements in visual quality and prompt fidelity.
Text-to-Image & Video Generation/Editing
Benchmarking and Mitigating Upstream LLM Prompting Dependencies in Text-to-Visual Generation
Modern text-to-image and text-to-video generation and editing methods rely heavily on external large language models (LLMs) for prompt rewriting, sub-prompt decomposition, and attribute filtering. Consequently, visual generation quality is bottlenecked by upstream LLM biases, template artifacts, and hallucinated attributes, frequently failing on ambiguous or in-the-wild user instructions unless manually tuned. Because no systematic study has evaluated visual generation frameworks across diverse upstream prompt generators and decomposition strategies, the sensitivity and failure modes of these pipelines remain unquantified.