OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Image & Video AI

Alignprop

Alignprop is an open-source Python research implementation that fine-tunes text-to-image diffusion models by backpropagating differentiable reward gradients through the denoising process. It updates low-rank adapter weights and uses gradient checkpointing to reduce memory demands.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Alignprop — OpenViST project profile in Image & Video AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Alignprop when you want a code-based research workflow for adapting a text-to-image diffusion model to differentiable reward objectives, rather than a hosted image-generation tool.

Who it is for

  • Researchers studying reward-based alignment of text-to-image diffusion models.
  • Developers experimenting with image-generation objectives such as aesthetics, semantic text-image alignment, compressibility, or object-count control.

Documented uses

  • Fine-tune a diffusion model using an aesthetic or HPSv2 reward training script.
  • Explore reward-gradient optimization for image-text alignment or other documented generation objectives.
  • Run training on a single GPU or use the documented Accelerate support for multi-GPU execution.
Before you choose

Requirements & limits

The repository is public and lists an MIT license. Its README documents a self-managed Conda and GPU training workflow.

  • The documented setup uses Python 3.10, Conda, and the repository requirements.
  • Training supports single-GPU use and describes Accelerate handling for multi-GPU execution.
  • The README reports four A100 GPUs with 40 GB of RAM for its experiments and suggests adjusting batch size or using K=1 when GPU memory is limited.
  • The repository includes specific aesthetic and HPSv2 scripts; other objectives are described as tested research targets.

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

Set up the Alignprop Conda environment

Create the documented Python environment and install the repository requirements.

Before you start

  • Conda
  • The Alignprop repository and its requirements.txt file

Steps

  1. Create an environment with Python 3.10 using conda create -n alignprop python=3.10.

  2. Activate it with conda activate alignprop.

  3. Install the dependencies with the command below.

    Documented command
    pip install -r requirements.txt

Expected outcome

A Conda environment named alignprop with the repository requirements installed.

Guide limitations

  • These instructions cover environment setup only, not verification or training execution.
Documentation 1 ↗
Source-based practical guide

Launch a documented reward-training script

Start one of the repository's documented aesthetic or HPSv2 training workflows.

Before you start

  • The Alignprop environment and repository requirements installed
  • GPU access for training

Steps

  1. For the aesthetic reward workflow, run the command below; the README also provides the command below for the K=1 variant.

    Documented command
    bash aesthetic.sh
    Documented command
    bash aesthetic_k1.sh

  2. For the HPSv2 reward workflow, run the command below; the README also provides the command below for the K=1 variant.

    Documented command
    bash hps.sh
    Documented command
    bash hps_k1.sh

  3. If GPU memory is a constraint, consider K=1 and adjust train_batch_size as the README suggests.

Expected outcome

The selected repository training script is launched for its documented reward objective.

Guide limitations

  • The instructions do not specify a universal GPU minimum; the authors report using four A100 GPUs with 40 GB of RAM for their experiments.
  • The README identifies num_epochs as a setting users may adjust for the aesthetic workflow.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
16 October 2026
Original website title
mihirp1998/AlignProp · Official homepage title · retained in its source language
How we review sourcesSuggest a correction

Related options

DDPO — The repository says its codebase is built directly on top of DDPO.