black-forest-labs

black-forest-labs / Self-Flow

Public

Code and website for Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis

41
2
100% credibility
Found Mar 04, 2026 at 41 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This repository provides inference code to generate thousands of 256x256 images using a pre-trained Self-Flow diffusion model on ImageNet classes for quality evaluation.

How It Works

1
📰 Discover Self-Flow

You hear about a cool tool from Black Forest Labs that creates realistic images of everyday things like animals, cars, and objects using smart AI.

2
📥 Grab the files

You download the simple package of files to your computer to get started with image creation.

3
🧠 Add the trained brain

You place a ready-made trained file into a special folder so the tool knows exactly how to draw lifelike pictures.

4
✨ Launch image generation

You start the generator, telling it to create thousands of new images, and feel the excitement as it works its magic.

5
📁 Watch images appear

A folder fills up with colorful 256x256 pixel images of random categories, ready for you to explore.

6
📊 Check quality

You use a helper tool to score how real and varied your new images look compared to real photos.

🏆 Celebrate great results

Your AI-generated images score high on realism tests, proving the tool's power for creative or research fun.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 41 to 41 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is Self-Flow?

Self-Flow is a Python-based GitHub code repository for running inference on a self-supervised flow matching model that generates 256x256 ImageNet images. Drop in a pretrained checkpoint, and use the CLI script with torchrun for multi-GPU sampling—generate 50k images in NPZ format for FID evaluation via the ADM suite. It handles SDE or ODE modes, classifier-free guidance, and saves PNGs alongside metrics-ready outputs.

Why is it gaining traction?

This GitHub Python code stands out for scalable multi-modal synthesis without paired data, delivering paper-level FID scores via simple commands like `torchrun sample.py --ckpt model.pt --num-fid-samples 50000`. Developers grab it for quick benchmarking against diffusion baselines, with built-in multi-GPU scaling and VAE decoding that plugs into standard eval pipelines—no fussing with custom setups.

Who should use this?

AI researchers replicating flow matching papers or tuning diffusion samplers on ImageNet. ML engineers evaluating generative models for multimodal apps, especially those needing 50k-sample FID runs on 8+ GPUs. Benchmarkers tired of wrangling OpenAI's guided-diffusion evals.

Verdict

Solid starter for Self-Flow inference with a crisp README and working CLI, but at 41 stars and 1.0% credibility score, it's early-stage—expect tweaks for production. Grab the checkpoint and benchmark if you're in gen AI; skip unless ImageNet evals are your jam.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.