drdaxxy

Typical approaches to training, and sampling from, denoising diffusion models yield results whose per-item means match the initial input - i.e. zero when using i.i.d. samples from a standard normal distribution. This has major implications for what outputs can be obtained from popular text-to-image generative models, see e.g. https://twitter.com/apeoffire/status/1624884816851206145 and https://www.crosslabs.org/blog/diffusion-with-offset-noise.

It also means we can reliably produce dark, bright, or tinted images by shifting the input to a desired color.

Now, I was curious what would happen if I made Stable Diffusion denoise an "impossible" image whose mean color exceeds the [0,1] valid RGB range:

init_latent = vae_encode(tensor([1.5, 1.5, 1.5])[None,:,None,None].tile(1,1,512,512)) + sigma_max * randn(1,4,64,64)

drdaxxy

{
   "diffusion_out":[
      "model.diffusion_model.output_blocks.6.0.in_layers.0.bias",
      "model.diffusion_model.output_blocks.6.0.in_layers.0.weight",
      "model.diffusion_model.output_blocks.6.0.out_layers.0.bias",
      "model.diffusion_model.output_blocks.6.0.out_layers.0.weight",
      "model.diffusion_model.output_blocks.7.0.in_layers.0.bias",
      "model.diffusion_model.output_blocks.7.0.in_layers.0.weight",
      "model.diffusion_model.output_blocks.7.0.out_layers.0.bias",
      "model.diffusion_model.output_blocks.7.0.out_layers.0.weight",

drdaxxy

"Antiprompts" in Craiyon / DALL-E Mega

Current image generation models use a technique called classifier-free guidance[^1] or super conditioning[^2]: At each step, they draw not just what fits the prompt best, but what sets images fitting that description apart from ones that reach similar intermediate results so far but might fit a totally different caption in the end.

To do that, they predict what they'd do next if the prompt was blank and decide based on the difference to what they'd do for the real prompt alone.

There ain't no rule that says we can't change that blank "antiprompt" to something meaningful, though. For Stable Diffusion, others have shown what replacing it can do (to try it, see e.g. AUTOMATIC1111's UI).

drdaxxy

drdaxxy

Grabbing TechnologyGuide forums via Tapatalk

Most of this applies to any forum using a (sufficiently up-to-date) version of the Tapatalk plugin. Some things may be specific to XenForo 1 or TechnologyGuide's setup. Users and forum administrators may also want to export private data (PMs, restricted forums, edit history...) which is out of scope here.

• Error responses: 200 OK with body like:

result = /* repsonse body */ {
  result: false,
  result_text: "Need valid forum id!",
  error: "Need valid forum id!"

drdaxxy

/* yolo */

.App {
  display: flex;
  flex-direction: column;
  align-items: center;
  justify-content: center;
}

form {

drdaxxy

'use strict';

// finds all used BGM IDs in Iruna Online
//
// to get tracks, you can rewrite an ID list into URLs:
//   <realm resource base>/bgm/<id>.zip
// e.g. 0 => http://rs-jp.iruna.jp/release/bgm/0.zip
// then download and extract with ZIP_PASSWORD
//
// only difference between realms seems to be updates

drdaxxy

struct {
    // Unknown file header fields have no effect on the animation
    uint unk0;
    uint unk1;
    uint unk2;
    uint unk3;
    uint unk4;
    uint unk5;
    uint unk6;
    uint unk7;

drdaxxy

// THIS FILE WILL NO LONGER BE UPDATED
// This code now lives at
// https://github.com/drdaxxy/ffxvDitherPatch/blob/master/Craf/CrafArchive.cs
// --------------------------------------------------------------------------


// THIS FILE WILL NO LONGER BE UPDATED


// Copyright 2018 Niklas K.

drdaxxy

#!/usr/bin/env python2

# requires lxml, requests, cssselect, pillow

import sys
import os
from io import BytesIO
from lxml import html
import requests
from PIL import Image