Practical overview
When it fits
Use the demo to try image generation interactively; use the repository when you need the released model code, weights, notebooks or documented training-data format for technical work.
Who it is for
- People who want to try text-to-image generation through the linked interactive demo.
- Developers seeking model code and checkpoints for self-hosted inference.
- Researchers exploring autoregressive image-generation methods or preparing training data.
Documented uses
- Generate images from text prompts using the linked demo or repository inference code.
- Inspect Infinity-2B and Infinity-8B inference with the provided interactive notebooks.
- Prepare a JSONL image-caption dataset in the documented format for training.
Requirements & limits
The repository links an interactive web demo and publicly provides code and model checkpoints. The project is listed under the MIT license.
- The README specifies PyTorch 2.5.1 or later for FlexAttention and installation of additional packages from requirements.txt.
- The listed setup also requires Infinity VAE and transformer weights, plus FLAN-T5-XL.
- The README identifies Infinity-8B checkpoints at 512×512 and 1024×1024.
- Training data uses JSONL files whose names encode aspect-ratio templates and example counts. The README links a 10-image toy dataset as a preparation reference.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the documented environment for Infinity
Set up the listed software and model weights before using the repository’s inference code.
Before you start
- A Python environment for the repository.
- PyTorch 2.5.1 or later for FlexAttention.
- The repository’s requirements.txt file.
- Infinity VAE and transformer weights from the linked Hugging Face collection.
- The FLAN-T5-XL tokenizer and model.
Steps
Install the other Python packages with pip3 install -r requirements.txt, as directed by the README.
Download the VAE and transformer weights from the FoundationVision Infinity Hugging Face collection.
Load FLAN-T5-XL with T5Tokenizer and T5ForConditionalGeneration using the model identifier google/flan-t5-xl.
Expected outcome
The README’s listed software and model dependencies are prepared for the next inference step.
Guide limitations
- The cited installation instructions do not document the complete inference command or a successful run.
- The listed PyTorch requirement is specifically associated with using FlexAttention to speed up training.
Prepare a training dataset in the documented layout
Organize text-to-image training examples into the JSONL structure described by the repository.
Before you start
- Image files accessible at paths recorded in the dataset entries.
- JSONL files named with an aspect-ratio template and example count, in the format [h_div_w_template]_[num_examples].jsonl.
- Each required record must include image_path, h_div_w, long_caption and long_caption_type.
Steps
Create a JSONL file for each aspect-ratio template, including the template ratio and number of examples in its filename.
Add one JSON item per line, supplying the image path, height-to-width ratio, long caption and long-caption type.
Optionally include text and short_caption_type for short captions.
Refer to the repository’s 10-image toy dataset for an example of data preparation.
Expected outcome
A dataset arranged in the format the README says its iterable dataset supports for training.
Guide limitations
- The source describes the data layout but this guide does not cover launching or completing a training run.
- The README says example counts for each aspect-ratio template must be specified in the filename.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- FoundationVision/Infinity · Official homepage title · retained in its source language