--- language: - en license: apache-2.0 library_name: diffusers pipeline_tag: text-to-video base_model: JaydenLu666/Reward-Forcing-T2V-1.3B base_model_relation: finetune datasets: - minar09/InfiniFlow-Bench tags: - text-to-video - video-generation - diffusion - autoregressive - long-horizon - long-video - fixed-camera - steady-forcing - nature-flow - static-view - spatial-persistence - motion-continuity - nature-video - fluid-dynamics - background-anchoring - temporal-consistency - dual-memory - wan2.1 - reward-forcing - arxiv:2606.14732 instance_prompt: >- Steady-Forcing: Balancing, Spatial Persistence, Motion Continuity, Long-Horizon, Nature Video Diffusion, drift-stagnation, nature-flow, fixed-camera, static nature widget: - text: >- A beautiful South Korean empty sea beach scene recorded by a completely fixed, static, tripod mounted camera. The camera is not seen, it does not move, tilt, pan, or zoom at any point. The sandy shoreline stretches across the frame, with gentle waves rolling in from the horizon and softly breaking against the beach in a continuous rhythm. The motion of the tides is natural and consistent, creating subtle ripples and foamy edges as the water meets the sand. The beach itself remains perfectly static, with smooth sand, scattered shells, and distant rocks unmoving in the background. The atmosphere is stunning, as the sun hangs low near the horizon, casting warm golden and orange tones across the sky and reflecting softly on the water's surface. The distant sea remains calm, with no flicker, jumps, resets, or artificial distortions, while the tide flows seamlessly in one constant direction. The scene unfolds in real time over a long duration, emphasizing the physical dynamics of the ocean waves and the tranquil beauty of the sunset. The video maintains temporal continuity across all frames, showing the uninterrupted progression of tides softly hitting the sandy beach without cinematic exaggeration or dramatic effects. [60s] parameters: negative_prompt: >- 色调艳丽,过曝,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量,JPEG压缩残留,丑陋的,残缺的,多余的手指,画得不好的手部,画得不好的脸部,畸形的,毁容的,形态畸形的肢体,手指融合,静止不动的画面,杂乱的背景,三条腿,背景人很多,倒着走, solid water, metallic water, water stagnation, color drift, water flow drift, water color drift, water surface drift, scene drift, background drift, frozen motion, camera movement, zooming, panning, camera drift, visual artifacts, camera, tripod, ground artifacts, anomalous textures, unrealistic round shaped pattern, localized distortions, unnatural water, unnatural waves, unnatural flow, unnatural motion, unnatural physics, unnatural dynamics, unnatural fluidity, unnatural surface, unnatural reflections, unnatural refractions, unnatural transparency, unnatural opacity, unnatural viscosity, unnatural turbulence, unnatural splashes, unnatural ripples, unnatural foam, unnatural spray, unnatural mist, unnatural droplets, human, animal, repetitive round textures, pond effects, low dynamic degree, unnatural color, unnatural lighting, unnatural shadows, unnatural highlights, unnatural contrast, unnatural saturation, unnatural hue, unnatural brightness, unnatural darkness, unnatural exposure, unnatural noise, unnatural grain, unnatural blur, unnatural sharpness, unnatural clarity, unnatural detail, unnatural texture, unnatural pattern, unnatural composition output: url: assets/A_beautiful_South_Korean_empty_sea_beach_scene_recorded_by_a_completely_fixed,_static,_tripod_mounte-0-1779332549.1318147.mp4 co2_eq_emissions: emissions: 96000 source: Estimated — 67 training hours on 8x NVIDIA A100 80GB GPUs at ~400W each, Korean grid (~0.45 kg CO2/kWh) training_type: fine-tuning geographical_location: South Korea hardware_used: 8x NVIDIA A100 80GB GPUs model-index: - name: Steady-Forcing-T2V-1.3B results: - task: type: text-to-video name: Text-to-Video Generation benchmark: name: VBench prompts: prompts/prompts5s.txt split: eval (5s) metrics: - name: Background Consistency Score (BCS) type: background-consistency-score value: 98.06 - name: Motion Smoothness Score (MSS) type: motion-smoothness-score value: 99.00 - name: Temporal Flicker Index (TFI) type: temporal-flicker-index value: 98.53 - name: Imaging Quality Score (IQS) type: imaging-quality-score value: 69.87 ---

Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion

arXiv GitHub Hugging Face Project Page


## ✍️ Authors [Matiur Rahman Minar](https://minar09.github.io/)1, [Seunghun Oh](https://owinhun.github.io/)2, [Ganghyeon Jeong](https://github.com/Jugahy)2, [Unsang Park](https://cviplab.sogang.ac.kr/cviplab/1249.html)1,2 1[Department of Computer Science and Engineering, Sogang University](https://ecs.sogang.ac.kr/ecs/index_new.html)   2[Department of Artificial Intelligence, Sogang University](https://ai.sogang.ac.kr/ai/index_new.html) ## 🚀 Progress - [x] 📝 Technical Report / Paper - [x] 🌐 Project Homepage - [x] 💻 Training & Inference Code - [x] 🤗 Pretrained Model: T2V-1.3B ## 🎯 Overview Steady-Forcing produces long-horizon nature video rollouts from a fixed-camera view. It decouples spatial persistence from motion continuity via a structural dual-memory protocol. This enables stable backgrounds and sustained fluid motion.
> **TL;DR**: We propose a dual-memory framework that balances stability and motion to sustain high background persistence and continuous fluid dynamics over multi-minute horizons for fixed-camera nature video generation. ## 🎬 Demo https://minar09.github.io/steadyforcing/ --- ## 🔧 Requirements | Requirement | Specification | |---|---| | GPU | NVIDIA GPU with ≥ 24 GB VRAM (tested on A100 80 GB) | | OS | Linux | | Python | 3.10 | | Training Setup | 8 × A100 GPUs (for full training run) | > Other hardware configurations may work but have not been tested. --- ## 🛠️ Installation Clone the repository and set up the environment in one step: ```bash git clone https://github.com/minar09/steady-forcing.git cd steady-forcing bash setup_env.sh ``` This script creates a Python 3.10 environment, installs all dependencies from `requirements.txt`, and downloads the required base models. Alternatively, a `Dockerfile` is provided for containerized setups: ```bash docker build -t steady-forcing . ``` --- ## 📦 Pretrained Checkpoints ### Download ```bash huggingface-cli download minar09/Steady-Forcing-T2V-1.3B --local-dir ./ckpt ``` Or using the Python API: ```python from huggingface_hub import snapshot_download snapshot_download(repo_id="minar09/Steady-Forcing-T2V-1.3B", local_dir="./ckpt") ``` > **Note:** The training algorithm is **data-free distillation** — no video dataset is required. ### File Structure After downloading, organize your working directory as follows: ``` steady-forcing/ ├── prompts/ # Example text prompts ├── configs/ # Model and training configs ├── pipeline/ # Inference pipeline ├── trainer/ # Training modules ├── demo_utils/ # Demo helper utilities ├── scripts/ # Utility scripts ├── templates/ # Prompt templates ├── ckpt/ │ └── steady-forcing-t2v.pt # Main model checkpoint ├── inference.py ├── inference.sh ├── train.py ├── train.sh ├── demo.py └── setup.py ``` --- ## 🚀 Inference ### Quick Start ```bash bash inference.sh ``` ### Custom Prompt Inference ```python from pipeline import SteadyForcingPipeline pipe = SteadyForcingPipeline.from_pretrained("minar09/Steady-Forcing-T2V-1.3B") prompt = """A serene woodland stream scene recorded by a completely fixed, static, tripod mounted camera. A narrow stream of clear water flows continuously from the upper part of the frame toward the lower edge, winding gently between moss-covered rocks and grassy banks. [60s]""" negative_prompt = """solid water, metallic water, water stagnation, color drift, frozen motion, camera movement, zooming, panning, visual artifacts, unnatural water, unnatural waves, unnatural flow, unnatural motion, human, animal""" video = pipe( prompt=prompt, negative_prompt=negative_prompt, num_frames=480, # 60s at 8fps ).frames video[0].save("output.mp4") ``` ### Recommended Negative Prompt For best results, always include this negative prompt to suppress common failure modes: ``` 色调艳丽,过曝,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量, JPEG压缩残留,丑陋的,残缺的,多余的手指,画得不好的手部,画得不好的脸部,畸形的,毁容的, 形态畸形的肢体,手指融合,静止不动的画面,杂乱的背景,三条腿,背景人很多,倒着走, solid water, metallic water, water stagnation, color drift, water flow drift, water color drift, water surface drift, scene drift, background drift, frozen motion, camera movement, zooming, panning, camera drift, visual artifacts, camera, tripod, ground artifacts, anomalous textures, unrealistic round shaped pattern, localized distortions, unnatural water, unnatural waves, unnatural flow, unnatural motion, unnatural physics, unnatural dynamics, unnatural fluidity, unnatural surface, unnatural reflections, unnatural refractions, unnatural transparency, unnatural opacity, unnatural viscosity, unnatural turbulence, unnatural splashes, unnatural ripples, unnatural foam, unnatural spray, unnatural mist, unnatural droplets, human, animal, repetitive round textures, pond effects, low dynamic degree, unnatural color, unnatural lighting, unnatural shadows, unnatural highlights, unnatural contrast, unnatural saturation, unnatural hue, unnatural brightness, unnatural darkness, unnatural exposure, unnatural noise, unnatural grain, unnatural blur, unnatural sharpness, unnatural clarity, unnatural detail, unnatural texture, unnatural pattern, unnatural composition ``` --- ## 🏋️ Training ### Self-Forcing Training with DMD ```bash bash train.sh ``` > Training was completed in **under 67 hours on 8 × A100 GPUs**. > No video dataset is required — the method uses **data-free ODE distillation**. ### Running the Demo ```bash python demo.py ``` ### Testing Model Loading ```bash python test_loading.py ``` --- ## 🔑 Trigger Words Include these keywords in your prompts to activate model-specific conditioning: | Trigger Word | Purpose | |---|---| | `Steady-Forcing: Balancing` | Core method conditioning | | `Spatial Persistence` | Static background anchoring | | `Motion Continuity` | Sustained fluid dynamics | | `Long-Horizon` | Extended duration generation | | `Nature Video Diffusion` | Natural scene domain | | `drift-stagnation` | Suppresses visual drift artifacts | | `nature-flow` | Natural fluid flow conditioning | | `fixed-camera` | Fixed tripod camera constraint | | `static nature` | Static environmental anchoring | --- ## 📊 Results Quantitative and qualitative results are available in the [arXiv preprint](https://arxiv.org/abs/2606.14732). For visualizations and video comparisons, visit the [project page](https://minar09.github.io/steadyforcing/). --- ## 📄 Citation If you use this model or codebase in your research, please cite our paper: ```bibtex @article{minar2025steady, title={Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion}, author={Minar, Matiur Rahman and Oh, Seunghun and Jeong, Ganghyeon and Park, Unsang}, journal={arXiv preprint arXiv:2606.14732}, year={2026} } ``` --- ## 🤝 Acknowledgements This project builds on the open-source [Infinity-RoPE](https://github.com/yesiltepe-hidir/infinity-rope) and [Reward-Forcing](https://github.com/JaydenLyh/Reward-Forcing) implementations, and acknowledges related work in long-horizon video diffusion, motion continuity, and spatial persistence. We sincerely appreciate their efforts and thank them.