OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Identity Grpo

Identity Grpo is an open-source Python research project that uses human-feedback-driven reinforcement learning to improve the consistency of multiple human identities in generated videos. It pairs a preference-trained video reward model with Identity-GRPO, a method designed for multi-person consistency, and documents training and inference workflows using VACE.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Identity Grpo — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when you need a research codebase for exploring reinforcement-learning-based identity preservation in multi-person generated video, and can provide the models, dataset preparation, and configured local environment.

Who it is for

  • Video-generation researchers studying multi-person identity consistency.
  • Teams prototyping human-preference-based refinement of generated video.
  • Practitioners adapting the documented VACE workflow who can prepare datasets and configure model weights.

Documented uses

  • Refine generated videos to better preserve the identities of multiple people during dynamic interactions.
  • Prepare train and test CSV data for the repository’s VACE workflow.
  • Run documented inference with configured pretrained, reward-model, and LoRA weights.
  • Investigate a preference-trained identity reward model separately through the referenced inference script.
Before you choose

Requirements & limits

The public GitHub repository is listed under the MIT license. Its documented use requires local setup and model-weight downloads; no hosted generation service is described.

  • The documented environment specifies Python 3.10.16, Conda, package installation, and FlashAttention2.
  • The README names Wan2.1-VACE-1.3B-diffusers and Qwen2.5-VL-3B-Instruct, along with separate LoRA and identity-reward weights.
  • Training and inference depend on user-prepared CSV data and configured paths.
  • The project reports up to 18.9% improvement in human-consistency metrics over baselines; this is a reported experimental result, not a performance guarantee.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run the documented VACE inference workflow

Prepare the repository configuration and test data for the README’s VACE inference command.

Before you start

  • The repository and its Python 3.10.16 Conda environment, with packages installed and FlashAttention2 installed.
  • Wan2.1-VACE-1.3B-diffusers and Qwen2.5-VL-3B-Instruct pretrained models, plus the Identity-GRPO LoRA and identity-reward weights.
  • A test CSV at dataset/generated_img/test.csv and configuration paths set in config/dgx.py and config/base.py.

Steps

  1. Set the pretrained-model and reward-model paths in config/dgx.py.

  2. Set train.lora_path in config/base.py to outputs/identity_grpo/ckpt/vace/lora, or the configured fine-tuned GRPO weight path.

  3. Prepare test data in dataset/generated_img/test.csv.

  4. Run the documented Accelerate command: accelerate launch --config_file scripts/accelerate_configs/multi_gpu.yaml --num_processes=1 --main_process_port 29501 test_wan2_1_vace.py --config config/dgx.py:vace.

Expected outcome

The repository’s VACE test/inference script is launched with the configured model weights and test dataset.

Guide limitations

  • This is a local repository workflow, not a hosted service. The README requires model downloads, configured paths, and prepared data.
  • The README gives a single-process inference launch; it does not establish hardware compatibility or guarantee a successful run.
Documentation 1 ↗
Source-based practical guide

Set up the Identity-GRPO environment

Create the documented Conda environment and install the repository’s packages and FlashAttention2 dependency.

Before you start

  • Git, Conda, Python 3.10.16, and the specified FlashAttention2 wheel for CUDA 12 and PyTorch 2.6.

Steps

  1. Clone https://github.com/alibaba/identity-grpo.git and enter the identity_grpo directory.

  2. Create and activate the Conda environment with Python 3.10.16: conda create -n vace_grpo python=3.10.16, then conda activate vace_grpo.

  3. Install the repository package with pip install -e .

  4. Install FlashAttention2 using the wheel command shown in the README.

Expected outcome

The documented Python environment and repository package installation steps are completed; model downloads and data preparation remain separate stages.

Guide limitations

  • The README specifies a particular FlashAttention2 wheel filename and platform build; this guide does not imply compatibility with other environments.
  • Environment setup alone does not run training or inference.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
alibaba/identity-grpo · Official homepage title · retained in its source language
How we review sourcesSuggest a correction