Practical overview
When it fits
Use Vader when you want a research codebase for adapting video diffusion models through reward gradients and can manage model checkpoints, GPU resources and Python environments.
Who it is for
- Researchers studying reward-based alignment for video generation.
- Developers experimenting with fine-tuning text-to-video diffusion models.
- Teams comparing reward signals and DPO or DDPO baselines in video-generation research.
Documented uses
- Fine-tune supported video diffusion models using reward-model feedback.
- Explore video-text similarity or aesthetic alignment objectives.
- Compare reward-based alignment experiments with DPO and DDPO baselines.
Requirements & limits
The GitHub repository is public and identified as open source. The supplied repository metadata does not specify a license.
- The README recommends beginning with VADER-VideoCrafter.
- The documented VideoCrafter inference example reports 16 GB of GPU VRAM at batch size one with fp16; the Open-Sora 360p example reports 40 GB with bf16.
- Stable Video Diffusion image-to-video adaptation and movie-generation code are listed as incomplete.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the VADER-VideoCrafter environment
Create the documented Conda environment and install the VideoCrafter-side dependencies.
Before you start
- A checkout of the VADER repository, with the commands run from its root directory.
- Conda.
- The repository's VADER-VideoCrafter directory.
Steps
Change into VADER-VideoCrafter.
Create and activate a Conda environment with Python 3.10.
Install PyTorch 2.3.0, torchvision 0.18.0, torchaudio 2.3.0 and pytorch-cuda 12.1 from the PyTorch and NVIDIA channels.
Install xformers and the packages in requirements.txt.
Clone HPSv2, install it in editable mode, then return to the VADER-VideoCrafter directory.
Expected outcome
The README's stated VADER-VideoCrafter environment and HPSv2 dependency are installed.
Guide limitations
- This covers environment setup only, not training or inference.
- The README's commands specify particular package versions and CUDA 12.1.
Run the documented VideoCrafter text-to-video inference script
Use the repository's provided VideoCrafter text-to-video inference script.
Before you start
- The VADER repository and its VADER-VideoCrafter environment and dependencies.
- Accelerate configured with accelerate config.
- The pretrained VideoCrafter2 checkpoint; the README gives a path for manual placement if it is not downloaded automatically.
- The README reports a single GPU with 16 GB of VRAM for batch size one with fp16 mixed precision.
Steps
From the VADER repository root, run accelerate config to set accelerator settings.
Change into VADER-VideoCrafter.
Run sh scripts/run_text2video_inference.sh.
Expected outcome
The repository's documented text-to-video inference script is invoked for VideoCrafter.
Guide limitations
- The README's reported 16 GB GPU configuration uses val_batch_size=1 and fp16 mixed precision.
- If the base model does not download automatically, the README says to place model.ckpt at VADER/VADER-VideoCrafter/checkpoints/base_512_v2/model.ckpt.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- mihirp1998/VADER · Official homepage title · retained in its source language