Telegram Logo Generator
A 4.9 million parameter neural network trained by KreatixLabs to generate exactly one image: the Telegram logo.
A highly specialized model, to put it generously.
Model Details
| Property | Value |
|---|---|
| Resolution | 1280 × 1280 |
| Parameters | 4,915,200 |
| Architecture | Trainable RGB pixel tensor |
| Format | Safetensors |
| Framework | PyTorch |
| Output diversity | None, by design |
How It Works
The model learns individual RGB pixel values through gradient descent, reconstructing the Telegram logo from its trained weights.
It is not a diffusion model and does not process text prompts.
Whether you request a landscape, a portrait, or the Telegram logo, the result remains remarkably consistent.
Training Results
- Mean absolute error: 3.19/255
- Exact pixel matches: 58.74%
- Resolution: 1280 × 1280
- Visual quality: Close to the original
Pixel-perfect reconstruction was not the objective, although the model made a respectable attempt.
Usage
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from PIL import Image
import numpy as np
path = hf_hub_download(
"KreatixLabs/telegram-logo-generator",
"telegram_model.safetensors"
)
weights = load_file(path)
pixels = weights["pixels"].clamp(0, 1)
image = Image.fromarray(
(pixels.numpy() * 255).round().astype(np.uint8)
)
image.save("telegram.png")
Limitations
- Generates only the Telegram logo.
- Does not support text conditioning.
- Not compatible with standard Stable Diffusion checkpoint loaders.
- Requires approximately 19 MB of weights to reconstruct an image that could be stored much more efficiently.
Efficiency was clearly not the primary research objective.
Source
Training image: https://cdn.kreatix.dev/file/Kreatix-pub/telegram-6896827_1280.png
Developed by KreatixLabs as an experimental single-image neural reconstruction project.
This project is not affiliated with Telegram. The Telegram logo remains a trademark of its respective owner.