DEV Community: norflin

Ready to use Google Colab to finetune SDXL + T2I Sketch Adapter 🤖🤯

norflin — Fri, 03 Nov 2023 13:54:13 +0000

https://github.com/norflin321/ml/blob/main/SDXL_autotrain_dreambooth.ipynb - open this colab, put your 1024x1024 images inside "./images" folder, and run all cells, after ~20 min it "pytorch_lora_weights.safetensors" file is ready to download from google colab.
https://github.com/norflin321/ml/blob/main/SDXL_sketch2img.ipynb - open this colab, put "pytorch_lora_weights.safetensors" file inside "./finetuned" folder, run all cells, first run will take ~5min, then 40 seconds to generate an image with SDXL finetuned on your images + your sketch image + your prompt

Good, fast, free, predictable results! 🤖🔥🤗

I made Software News Wallpaper in pure GO!

norflin — Mon, 19 Jun 2023 19:10:33 +0000

check out my new design for only software news wallpaper (no politics, no ads)
https://github.com/norflin321/os-pulse 😁

NVIDIA GPUs Performace Comparison (value per 1$)

norflin — Tue, 09 May 2023 00:04:51 +0000

Data Source
Code
Result

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import dataframe_image as dfi

plt.rcParams["figure.dpi"] = 300
plt.rcParams["savefig.dpi"] = 300
plt.rcParams["figure.figsize"] = [12, 5]

df = pd.DataFrame()
df['name'] = ['GeForce RTX 4090', 'GeForce RTX 4080', 'GeForce RTX 4070 Ti', 'GeForce RTX 4070', 'GeForce RTX 3090 Ti', 'GeForce RTX 3090', 'GeForce RTX 3080 Ti', 'GeForce RTX 3080', 'GeForce RTX 3070 Ti', 'GeForce RTX 3070', 'GeForce RTX 3060 Ti', 'GeForce RTX 3060']
df['release_date'] = ['October 12, 2022', 'November 16, 2022', 'January 5, 2023', 'April 13, 2023', 'March 29, 2022', 'September 24, 2020', 'June 3, 2021', 'January 27, 2022', 'June 10, 2021', 'October 29, 2020', 'October 27, 2022', 'September 1, 2021']
df['price'] = [2007.58, 1441.61, 1226.36, 875.47, 1368.88, 1328.24, 1177.12, 900.77, 910.68, 591.89, 610.55, 511.45]
df['tdp_watts'] = [450, 320, 285, 200, 450, 350, 350, 350, 290, 220, 200, 170]
df['ray_traycing_tflops'] = [191, 112.7, 92.7, None, 79.9, 71.1, 68.2, 61.3, 43.5, 40.6, 32.4, 25]
df['tensor_compute_tflops'] = [292, 172, 142, None, 269.1, 235.08, 228.6, 180.6, 154.8, 141.31, 109.7, 75.7]
df['half_precision_tflops'] = [73.1, 43.0, 35.5, 22.6, 33.5, 29.38, 28.06, 22.6, 19.35, 17.66, 13.70, 9.46]
df['double_precision_tflops'] = [1.142, 0.672, 0.554, 0.353, 0.524, 0.459, 0.438, 0.353, 0.302, 0.276, 0.214, 0.148]
df['single_precision_tflops'] = [73.1, 43.0, 35.5, 22.6, 33.5, 29.28, 28.57, 22.6, 19.35, 17.66, 13.72, 9.46]
df['fillrate_gts'] = [1290.2, 761.5, 626.4, 455.4, 524.2, 457.6, 438.5, 352.8, 302.36, 276.0, 214.3, 147.8]
df['fillrate_gps'] = [443.5, 280.6, 208.8, 158.4, 174.7, 156.2, 153.5, 131.0, 151.18, 144.0, 112.8, 63.4]
df['clock_speed_mhz'] = [2230, 2210, 2310, 1920, 1560, 1395, 1365, 1260, 1575, 1500, 1410, 1320]

for column in df.columns[3:]:
    df[column] = df['price'] / df[column]

df['tdp_watts'] = df['tdp_watts'] * -1
df = df.sort_values(by=['release_date'], ascending=True)
df['mean'] = df[df.columns[3:]].mean(axis=1)

def pretty_panda(x):
    return (df.style
                .background_gradient()
                .set_properties(**{'font-size': '10pt'})
                .format(precision=2)
                .set_caption("NVIDIA GPUs Performace Comparison (value per 1$)")
                .set_table_styles([
                    {'selector': 'thead', 'props': 'text-transform: uppercase;'},
                    {'selector': 'caption', 'props': 'font-size: 20px'}
                ]))

df = pretty_panda(df)
df

How to generate midjourney level images for free (forever) in cloud?

norflin — Fri, 31 Mar 2023 21:07:21 +0000

The answer is simple you can use my colab notebook 😄.

Сhange any hyperparameters if needed, then run 🤗.
It takes about 40 seconds to generate one image.

If you wanna change a model just select one from huggingface then paste its name into model_id variable.

Enjoy!

how to make any rounded shape way more sexy? (css only)

norflin — Wed, 29 Mar 2023 16:14:23 +0000

The problem with computer rounding is its unnaturalness.

If you build a rectangle in the graphical editor, and then add a fillet to it, a circle will simply be entered into its corner.

Because of this, a form arises that does not occur in the real world. To make the bend look natural, the line should change shape gradually.

Ultimately, the researchers found that users liked shapes with smooth curves more than those with sharp lines.

Sharp objects often signal danger. Slight curvature and bulges are characteristic of living organisms.

.squircle { aspect-ratio : 1/1; background: inherit; position: relative; border-radius: 20%; }
.squircle::before, .squircle::after { content: ""; position: absolute; z-index: -1; background: inherit; }
.squircle::before { border-radius: 2%/30%; top: 15%; bottom: 15%; right: -1%; left: -1%; }
.squircle::after { border-radius: 30%/2%; left: 15%; right: 15%; top: -1%; bottom: -1%; }

Just add .squircle class to your element and set width and background of the element.
the only limit is that the element must have a 1/1 ratio

stop wasting time on meaningless news!

norflin — Fri, 03 Feb 2023 21:36:24 +0000

Do you often find yourself distracted from productive work, by attention sucking and meaningless news or propaganda?
Stop it!
You can stay inform about open source projects and important news enough.
By reading only Github Trends and HackerNews.
But, wouldn't it be perfect if everything you needed was in one place?
Yes, it would.
Check this out!

make your vim editing faster!

norflin — Mon, 23 May 2022 18:48:15 +0000

i use vim every day and the best productivity improvement that happened to me lately is that i found this keyboard setting in macos, highly recommend everyone who likes to work with code quickly, set these sliders to the maximum!