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

Scalekv

ScaleKV is a Python research implementation of scale-aware transformer KV-cache compression for visual autoregressive image generation. It adapts cache allocation by transformer-layer role and prediction scale, and provides inference and sampling scripts for Infinity-2B and Infinity-8B.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Scalekv — 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 ScaleKV when working with the documented Infinity visual autoregressive models and exploring how adaptive KV-cache allocation affects inference memory and image samples.

Who it is for

  • Researchers studying visual autoregressive image generation and inference memory use.
  • Practitioners running the documented Infinity-2B or Infinity-8B workflows.
  • Teams comparing compressed and baseline image samples with the repository’s evaluation scripts.

Documented uses

  • Run Infinity-2B or Infinity-8B image inference with ScaleKV-compressed KV caches.
  • Sample images from COCO captions with compressed or baseline Infinity workflows.
  • Compare generated sample sets using PSNR, LPIPS and FID metrics.
Before you choose

Requirements & limits

The GitHub repository is public and identified as MIT-licensed. Its documented workflows use Python and Hugging Face model checkpoints.

  • The repository demonstrates Infinity-2B and Infinity-8B; using another visual autoregressive model is experimental and not shown in the implementation.
  • The README reports Infinity-8B memory use of 85 GB reduced to 8.5 GB, with GenEval at 0.79 and DPG changing from 86.61 to 86.49; these are project-reported results.
  • The documented setup requires installing repository dependencies and downloading model weights before using the inference scripts.
  • The repository includes separate baseline and compressed sampling scripts, followed by metric computation for comparing sample directories.

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 ScaleKV checkpoints for Infinity inference

Install the repository requirements and download the documented FLAN-T5-XL and Infinity-2B or Infinity-8B checkpoints.

Before you start

  • A Python environment with pip
  • The ScaleKV repository
  • Hugging Face Hub access for the documented checkpoint downloads

Steps

  1. From the repository, install dependencies with the command below.

    Documented command
    pip install -r requirements.txt

  2. Install or update the Hugging Face Hub client with the command below.

    Documented command
    pip install -U huggingface_hub

  3. Download google/flan-t5-xl to ./weights/flan-t5-xl using the documented huggingface-cli download command.

  4. Download the Infinity-2B or Infinity-8B checkpoint files using the repository’s respective documented download commands.

Expected outcome

The documented dependencies and selected model checkpoints are installed in the paths used by the repository’s inference workflow.

Guide limitations

  • This covers setup only; it does not establish that inference has run successfully.
  • The repository documents checkpoints for Infinity-2B and Infinity-8B; support for other models is not shown. The exact Infinity checkpoint commands are in the README.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
StargazerX0/ScaleKV · Official homepage title · retained in its source language
How we review sourcesSuggest a correction