A free exercise dataset with images: 708 exercises at 512 px

Community Article
Published November 15, 2026

Free exercise data usually comes with a catch. Some of it has no pictures, some has pictures the size of a thumbnail, and in some the pictures aren't yours to use. We wanted the free version of RepDB to be something you can actually build an app on. So it's the whole catalogue, with proper images, and it's here on Hugging Face as RepDB/exercise-dataset.

Barbell back squat, flat illustrationBarbell bench press, flat illustrationBarbell deadlift, flat illustrationPull-up, flat illustration
Push-up, flat illustrationLunge, flat illustrationRomanian deadlift, flat illustrationPlank, flat illustration

All 708 exercises, not a sample

The free tier has every exercise in RepDB: 708 as of the November 14 release, the same list as the paid bundle. That's:

  • 581 strength;
  • 88 stretching;
  • 19 cardio;
  • 16 olympic lifting;
  • 4 plyometric.

217 of them need no equipment. We update the free dataset with every release.

512 px, not 180

Each image is a 512×512 WebP. The most-starred exercise dataset on GitHub, hasaneyldrm/exercises-dataset with over 23,000 stars, ships its pictures and GIFs at 180×180, and so does the free ExerciseDB API. 512 px is about eight times the pixels. It's sharp in a list and big enough for an exercise card on a phone.

Moving exercises get two images, the start and the peak of the rep. Holds and stretches get one. That's 1,231 exercise images in total, plus 64 equipment and 29 muscle illustrations.

One style all the way through

Every illustration uses the same flat style, the same character and the same solid background. That sounds minor until you mix sources: a list where every third exercise looks different reads as unfinished. Here it looks like one set because it is one set.

What each exercise has

  • names, a short description, step-by-step instructions and tips in English, German and Spanish;
  • primary and secondary muscles;
  • equipment, body part and category;
  • mechanic, force type and difficulty;
  • goals and a MET value;
  • flags for unilateral and bodyweight moves.

exercises.json has everything, in all three languages. exercises.csv has one row per exercise in English, and it's what the dataset viewer shows.

Load it

from datasets import load_dataset

ds = load_dataset("RepDB/exercise-dataset", split="train")
print(ds[0]["name_en"], ds[0]["primary_muscles"])

The full JSON and all the images:

import json
from huggingface_hub import snapshot_download

root = snapshot_download(
    "RepDB/exercise-dataset",
    repo_type="dataset",
    allow_patterns=["exercises.json", "images/flat/*"],
)
with open(f"{root}/exercises.json", encoding="utf-8") as f:
    data = json.load(f)
print(data["count"], "exercises")

The same data is on npm as @repdb/exercises, on GitHub, and browsable at exercise-dataset.com.

Want GIFs? Two images are enough

We went through the GIFs in that 23,000-star repo frame by frame. Each one is two poses with a crossfade between them. The start is held for a second, then five blended frames, then the end for a second, then back.

Our free tier has the same two poses for every moving exercise, at 512 px. So you can make the same kind of GIF for your app in a few lines, with root from the snippet above:

from PIL import Image

def crossfade_gif(start_path, peak_path, out_path, steps=5, hold=1000, step=100):
    a = Image.open(start_path).convert("RGB")
    b = Image.open(peak_path).convert("RGB")
    blends = [Image.blend(a, b, i / (steps + 1)) for i in range(1, steps + 1)]
    frames = [a, *blends, b, *reversed(blends)]
    durations = [hold, *[step] * steps, hold, *[step] * steps]
    frames[0].save(out_path, save_all=True, append_images=frames[1:],
                   duration=durations, loop=0)

crossfade_gif(f"{root}/images/flat/squat-start.webp",
              f"{root}/images/flat/squat-peak.webp", "squat.gif")

That gives you a 512×512 GIF with the same timing. It's still a crossfade, not real movement. Real movement is what our paid animations are for.

The licence, plainly

It's free for personal and commercial use inside apps, websites and research projects. The one condition is a visible credit, "Exercise data by RepDB (repdb.co)", in your credits screen, README or footer. The link can be nofollow.

It isn't open data, though:

  • You can't republish it as a dataset or an API.
  • The images can't be used to train or condition generative models.

If you need something fully open, free-exercise-db (photos, Unlicense) is the honest pick.

Other free options

RepDB free hasaneyldrm/exercises-dataset ExerciseDB free API free-exercise-db wger
Exercises 708, the full catalogue 1,324 1,500 800+ 973 entries
Images 1,231 flat WebP, 512×512 a JPG and a crossfade GIF each, 180×180 one GIF each, 180×180 JPG photos of real people 394 images in total
Can you use the images in your app? yes, with a credit not from the repo; you need your own licence from Gym visual check the API's terms yes (Unlicense) check each image's licence
How you get it files: Hugging Face, GitHub, npm, Kaggle files on GitHub API, no key files on GitHub public API

Checked on November 15, 2026, on each project's own page, repo or API.

The star count on that GitHub repo hides one thing. Its MIT licence covers the code and the text only. Its own LICENSE says the images and GIFs belong to Gym visual and are included at 180×180 with Gym visual's permission. It also says that cloning the repo doesn't give you a licence to use them.

When you need animations

The free tier is still images only. The paid tiers add more:

  • Stills ($299): a second, more detailed style on transparent backgrounds, with the working muscles highlighted, at 1024 px.
  • Standard ($499): everything in Stills, plus 551 looping animations. The other 157 exercises are holds and stretches, so they stay as stills.

Both are one payment. The gallery shows 30 of the clips.

Questions about the data or the licence: ask in the dataset's Community tab.

Community

Sign up or log in to comment