You’ll never believe this, but…

That Amazon package that went from California to Illinois, had no tracking updates for five days, then appeared back in California? Supposedly left California early Monday morning and hasn’t been spotted since. It’s “still on the way”, allegedly by Friday, but has reached the point where Amazon is now offering me a refund.

Since I don’t actually need it for anything soon, I’m just going to see what happens. Will it suddenly acquire a new tracking number or shipping company, as they’ve done before, or will it just show up a few months from now, as they’ve also done before.

The Flying Sister

Not to be confused with The Flying Nun. I was expecting my sister to fly into town on Wednesday. Monday afternoon she called to say she’d missed a connection on one of her regular business flights, and if she was going to have to Zoom into a meeting, she could do that from my house. So she was going to come to my place early.

Then she (and 20 other planeloads of people) got stuck on the tarmac at O’Hare and missed that connection, with no later flights to Dayton. Fortunately she lives in Chicago, so she wasn’t just stuck in an airport overnight. Although it took them three hours to find her bags…

So close…

They almost did the meme:

Waifu-a-go-go

[trivia: the Hollywood nightclub “Whisky a Go Go” was named after the first French disco club, which was named after the British movie “Whisky Galore!” (a-gogo being Frogspeak for “galore”)]

The dynamic wildcarding is shaping up nicely (although I need to split it up into categorized sets, and generate a wider variety now that I’ve got the prompting down), and I was in the mood to generate a big batch of pinup gals, but it just takes too damn long, and I can’t fire up a game on the big PC while all its VRAM is being consumed fabricating imaginary T&A. Belatedly it occurred to me that if I’m going to do a separate refine/upscale pass on the good ones anyway, why not do the bulk generation at a lower resolution?

Instead of the nearly-16x9 resolution of 1728x960 upscaled 2.25x to 3888x2160, I dropped it to 1024x576, which can be upscaled 3.75x to exactly the monitor’s 3840x2160 resolution. That cuts the basic generation time from 90 seconds to 35, and if I also give up on using Heun++ 2/Beta for the upscaling (honestly, the improvements are small, and it changes significant details often enough to force me to retry at least once), the refine/upscale time drops from 33 minutes to 10. That makes it less annoying to ask for several hundred per batch.

I did get some odd skin texturing on the first upscaled image, so I tried switching to a different upscaler. That changed a bunch of details, so now I’ve downloaded new upscalers to try. TL/DR, the more you upscale, the more of your details are created by the upscaler.

On a related note, file under peculiar that in MacOS 15.x, Apple decided to strictly enforce limits on how often you can rotate wallpaper. It used to be that the GUI gave you a limited selection but you could just overwrite that with an AppleScript one-liner. Nope, all gone; now you’re only permitted to have the image change every 5 seconds, 1 minute, 5 minutes, 15 minutes, 30 minutes, 1 hour, or 1 day. No other intervals are considered reasonable. Apple knows you don’t need this.

No doubt the QA team that used to test this stuff was axed to fund the newly-released Liquid Ass GUI that unifies all Apple platforms in a pit of translucent suck.

(just as I’ve settled on “genai” as shorthand to refer to the output of LLMs and diffusion models, I’ve decided I need a short, punchy term for the process of building prompts, selecting models and LoRA, and iterating on the results; I could, for instance, shorten the phrase “Fabricating Pictures” to, say, “fapping”…)

Despite its many flaws, I really think this is the most promising model to play with, because the bolted-on LLM parser significantly improves your ability to lay out an image and add text.

This image came out almost exactly as requested. I ran it multiple times to select the best variation on the theme (sexy pin-ups, size and style of dice, length of liquor shelf, facial expressions, contents of character sheets, etc), but they were all correct.

The exact prompt was:

A gritty, realistic fantasy RPG illustration set in a tavern, with a large central table around which are seated five bearded dwarves wearing leather armor, with axes slung on their back. On the table in front of each dwarf are RPG character sheets, polyhedral dice, and wooden tankards filled with ale. The wall behind them has shelves full of liquor bottles with rune-covered labels, and pin-up posters of sexy goblin girls. The center dwarf is speaking, and the speech bubble contains the words “Okay, now make a shaving throw…” in a handwriting font.

Naturally, the speech bubble works with my usual sort of imagery as well…

(I didn’t specify maid lingerie, but I did ask for “elf maidens”, so call it a lucky accident)

Dear Microsoft,

I just had to email a plain-text attachment as a ZIP file from a Mac to a Mac to keep Outlook from mangling UTF-8 into random garbage. Fix the little shit before you shove your “AI” into every app, m’kay?

Coding with ChatGPT…

End every request after the first with the following words. You’ll thank me later.

Please fix ONLY this issue, and write out the complete corrected program, without making ANY unrequested changes to the code.

This is pretty much the only way to get readable diffs of your iterations. Otherwise there will be random changes everywhere. Comments added/deleted, code collapsed onto one line or expanded to multiples, functions renamed or reordered, regressions created, etc, etc.

Final ChatGPT Deathmatch

This is version 18.5; ChatGPT was still offering me additional enhancements, but we’d far exceeded my patience, even without the four-hour delay while my free credits recharged. The last two revisions were fixes for bugs that only emerged in final QA, where ranking/flagging an image with a value that was currently filtered out of the display list double-advanced, skipping over a still-valid image. Since the whole point of the method is progressively assigning ranks to every image, skipping one was a nuisance.

The “.5” comes from me not wanting to make Yet Another ChatGPT Pass to fix two lines of Javascript that were easily located in the file. The LLM had argued its way into violating the spec at the last minute while fixing the final bug.

It’s GenAI code, so by definition it can’t be copyrighted; if you have any use for it, knock yourself out.

#!/usr/bin/env python3
import sys, os, threading
from pathlib import Path
from flask import Flask, send_file, request, jsonify, render_template_string, abort

app = Flask(__name__)

BASE_DIR = Path(sys.argv[1] if len(sys.argv) > 1 else os.getcwd()).resolve()
RANK_FILE = BASE_DIR / "_rank.txt"
state_lock = threading.Lock()
files = []
meta = {}

# ------------------------------------------------------------
# Load/save ranking state
# ------------------------------------------------------------
def load_state():
    global files, meta
    files = sorted(
        [f.name for f in BASE_DIR.iterdir() if f.suffix.lower() in (".png", ".jpg")],
        key=lambda fn: (BASE_DIR / fn).stat().st_mtime,
    )
    meta = {}
    if RANK_FILE.exists():
        for line in RANK_FILE.read_text().splitlines():
            parts = line.split("\t")
            if not parts:
                continue
            fname = parts[0]
            rank = int(parts[1]) if len(parts) > 1 and parts[1] else 0
            flags = set(parts[2].split(",")) if len(parts) > 2 and parts[2] else set()
            meta[fname] = {"rank": rank, "flags": flags}
    for f in files:
        if f not in meta:
            meta[f] = {"rank": 0, "flags": set()}


def save_state():
    with open(RANK_FILE, "w") as fp:
        for fname in files:
            entry = meta.get(fname, {"rank": 0, "flags": set()})
            flags_str = ",".join(sorted(entry["flags"])) if entry["flags"] else ""
            fp.write(f"{fname}\t{entry['rank']}\t{flags_str}\n")


# ------------------------------------------------------------
# Routes
# ------------------------------------------------------------
@app.route("/")
def index():
    return render_template_string(INDEX_HTML)


@app.route("/image/<path:fname>")
def get_image(fname):
    if not fname:
        abort(404)
    target = (BASE_DIR / fname).resolve()
    if not str(target).startswith(str(BASE_DIR.resolve())):
        abort(403)
    if not target.exists():
        abort(404)
    return send_file(str(target))


@app.route("/api/state")
def api_state():
    with state_lock:
        safe_meta = {
            fname: {
                "rank": entry.get("rank", 0),
                "flags": sorted(entry.get("flags", [])),  # sets → sorted lists
            }
            for fname, entry in meta.items()
        }
        return jsonify({"files": files, "meta": safe_meta})


@app.route("/api/update", methods=["POST"])
def api_update():
    data = request.json
    fname = data.get("file")
    if fname not in meta:
        abort(400)
    with state_lock:
        if "rank" in data:
            meta[fname]["rank"] = data["rank"]
        if "toggle_flag" in data:
            fl = data["toggle_flag"]
            if fl in meta[fname]["flags"]:
                meta[fname]["flags"].remove(fl)
            else:
                meta[fname]["flags"].add(fl)
        save_state()
    return jsonify(success=True)


@app.route("/api/reload", methods=["POST"])
def api_reload():
    with state_lock:
        load_state()
    return jsonify(success=True)


# ------------------------------------------------------------
# HTML/JS template
# ------------------------------------------------------------
INDEX_HTML = """
<!doctype html>
<html>
<head>
<meta charset="utf-8"/>
<title>Five-Star Deathmatch</title>
<style>
body { margin:0; font-family:sans-serif; display:flex; flex-direction:column; height:100vh; }
.topbar, .bottombar { background:#eee; padding:4px; display:flex; justify-content:space-between; font-size:20px; }
.central { position:relative; flex:1; background:#999; display:flex; justify-content:center; align-items:center; overflow:hidden; }
.central img { max-width:100%; max-height:100%; object-fit:contain; }
.central.zoom { overflow:auto; justify-content:flex-start; align-items:flex-start; }
.central.zoom img { max-width:none; max-height:none; width:auto; height:auto; display:block; }
#help { position:absolute; top:10px; right:10px; cursor:pointer; border:1px solid #666; border-radius:8px; margin:2px; padding:6px; background:#f9f9f9; text-align:center; }
#helpPanel { display:none; position:absolute; top:40px; right:10px; background:#fff; color:#000; font-size:16px; padding:10px 15px; border-radius:10px; box-shadow:0 2px 8px rgba(0,0,0,0.3); }
#helpPanel ul { margin:0; padding-left:20px; }
#helpPanel li { margin:4px 0; }
.bottombar { flex-direction:column; font-size:14px; }
.row { display:flex; flex:1; }
.cell { flex:1; border:1px solid #666; border-radius:8px; margin:2px; text-align:center; padding:6px; cursor:pointer; display:flex; flex-direction:column; justify-content:center; align-items:center; background:#f9f9f9; }
.cell.active { background:#ccc; }
</style>
</head>
<body>
<div class="topbar"><div id="filename"></div><div id="rankflags"></div><div id="pos"></div></div>
<div class="central" id="central">
  <img id="mainimg"/>
  <div id="help">❓</div>
  <div id="helpPanel"></div>
</div>
<div class="bottombar">
  <div class="row" id="rankrow"></div>
  <div class="row" id="flagrow"></div>
</div>
<script>
let state={files:[],meta:{}};
let index=0;
let filters={ranks:new Set(),flags:new Set()};
let zoom=false;
const mainimg=document.getElementById("mainimg");
const filenameDiv=document.getElementById("filename");
const rankflagsDiv=document.getElementById("rankflags");
const posDiv=document.getElementById("pos");
const central=document.getElementById("central");
const help=document.getElementById("help");
const helpPanel=document.getElementById("helpPanel");

function fetchState(){ fetch("/api/state").then(r=>r.json()).then(js=>{state=js; render(); buildBottom();}); }
function buildBottom(){
  const rr=document.getElementById("rankrow"); rr.innerHTML="";
  const ranks=[-1,0,1,2,3,4,5];
  let counts={}; for(let r of ranks) counts[r]=0;
  let total=0;
  for(let f of state.files){ let m=state.meta[f]||{rank:0,flags:[]}; if(m.rank>-1) total++; counts[m.rank]++; }
  for(let r of ranks){
    let icon=(r==-1?"❌":(r==0?"⚪️":"⭐️".repeat(r)));
    let pct=(r>=0 && total>0)?Math.round(100*counts[r]/total)+"%":"";
    const d=document.createElement("div"); d.className="cell"; if(filters.ranks.has(r)) d.classList.add("active");
    d.innerHTML="<div>"+icon+"</div><div>"+counts[r]+(pct?" ("+pct+")":"")+"</div>";
    d.onclick=()=>{ if(filters.ranks.has(r)) filters.ranks.delete(r); else filters.ranks.add(r); render(); buildBottom(); };
    rr.appendChild(d);
  }
  const fr=document.getElementById("flagrow"); fr.innerHTML="";
  const fls=["A","B","C","D","E","F"]; let flagCounts={}; for(let fl of fls) flagCounts[fl]=0;
  let unflagged=0;
  for(let f of state.files){ let m=state.meta[f]||{rank:0,flags:[]}; if(m.flags.length==0) unflagged++; for(let fl of m.flags) flagCounts[fl]++; }
  for(let fl of fls){
    const d=document.createElement("div"); d.className="cell"; if(filters.flags.has(fl)) d.classList.add("active");
    d.innerHTML="<div>"+fl+"</div><div>"+flagCounts[fl]+"</div>";
    d.onclick=()=>{ if(filters.flags.has(fl)) filters.flags.delete(fl); else filters.flags.add(fl); render(); buildBottom(); };
    fr.appendChild(d);
  }
  const d=document.createElement("div"); d.className="cell"; if(filters.flags.has("UNFLAG")) d.classList.add("active");
  d.innerHTML="<div>🚫</div><div>"+unflagged+"</div>";
  d.onclick=()=>{ if(filters.flags.has("UNFLAG")) filters.flags.delete("UNFLAG"); else filters.flags.add("UNFLAG"); render(); buildBottom(); };
  fr.appendChild(d);
}
function filteredFiles(){
  return state.files.filter(f=>{
    let m=state.meta[f]||{rank:0,flags:[]};
    if(filters.ranks.has(m.rank)) return false;
    for(let fl of m.flags){ if(filters.flags.has(fl)) return false; }
    if(m.flags.length==0 && filters.flags.has("UNFLAG")) return false;
    return true;
  });
}
function render(){
  let list=filteredFiles(); if(list.length==0){ mainimg.src=""; filenameDiv.textContent=""; rankflagsDiv.textContent=""; posDiv.textContent="0/0"; return; }
  if(index>=list.length) index=0;
  let fname=list[index]; mainimg.src="/image/"+encodeURIComponent(fname); filenameDiv.textContent=fname;
  let m=state.meta[fname]||{rank:0,flags:[]};
  let rankDisp=(m.rank==-1?"❌":(m.rank==0?"⚪️":"⭐️".repeat(m.rank)));
  rankflagsDiv.textContent=rankDisp+" "+m.flags.sort().join("");
  posDiv.textContent=(index+1)+" / "+list.length;
}
function nextValidIndex(oldfname){
  let list=filteredFiles();
  let i=list.indexOf(oldfname);
  if(i==-1){
    if(index>=list.length) index=list.length-1;
  } else {
    index=i + 1;
  }
}
function updateRank(r){
  let fname=filteredFiles()[index];
  fetch("/api/update",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({file:fname,rank:r})})
    .then(()=>fetch("/api/state"))
    .then(r=>r.json())
    .then(js=>{
      state=js;
      nextValidIndex(fname);
      render(); buildBottom();
    });
}
function toggleFlag(fl){
  let fname=filteredFiles()[index];
  fetch("/api/update",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({file:fname,toggle_flag:fl})})
    .then(()=>fetch("/api/state"))
    .then(r=>r.json())
    .then(js=>{
      state=js;
      render(); buildBottom();
    });
}
document.addEventListener("keydown",ev=>{
  let list=filteredFiles(); if(list.length==0) return;
  if(ev.key=="ArrowLeft"){ index=(index-1+list.length)%list.length; render(); }
  else if(ev.key=="ArrowRight"){ index=(index+1)%list.length; render(); }
  else if(ev.key=="ArrowUp"){ index=0; render(); }
  else if(ev.key=="ArrowDown"){ index=list.length-1; render(); }
  else if(ev.key=="x"||ev.key=="X"){ updateRank(-1); }
  else if(ev.key>="0"&&ev.key<="5"){ updateRank(parseInt(ev.key)); }
  else if("ABCDEF".includes(ev.key.toUpperCase())){ toggleFlag(ev.key.toUpperCase()); }
  else if(ev.key=="r"||ev.key=="R"){ filters={ranks:new Set(),flags:new Set()}; render(); buildBottom(); }
  else if(ev.key=="z"||ev.key=="Z"){ zoom=!zoom; if(zoom){ central.classList.add("zoom"); central.scrollTop=0; central.scrollLeft=0; } else { central.classList.remove("zoom"); } }
  else if(ev.key==" "){ fetch("/api/reload",{method:"POST"}).then(()=>fetchState()); }
});
help.onclick=()=>{ 
  helpPanel.style.display="block"; 
  helpPanel.innerHTML="<ul>"
    +"<li>← / → : Previous / Next image</li>"
    +"<li>↑ / ↓ : First / Last image</li>"
    +"<li>0–5 : Set rank</li>"
    +"<li>X : Rank -1 (❌)</li>"
    +"<li>A–F : Toggle flags</li>"
    +"<li>R : Reset filters</li>"
    +"<li>Z : Toggle zoom</li>"
    +"<li>Space : Reload state</li>"
    +"</ul>";
};
helpPanel.onclick=()=>{ helpPanel.style.display="none"; };
fetchState();
</script>
</body>
</html>
"""

# ------------------------------------------------------------
if __name__=="__main__":
    load_state()
    app.run("127.0.0.1",5000,debug=True)

The odd mix of inlining and formatted blocks is how it went from 700+ lines to 268: every revised version made new formatting decisions in unrelated parts of the file until I started adding The Magic Words. Areas being updated in a revision got clean formatting, because it showed them in the explanation of the change, while “unchanged” areas got compressed onto one line.

(classical reference)

Dear Apple,

Please provide fine-grained controls for disabling “data detector” overlays in text and images. It’s really annoying to look at a picture that clearly contains no text at all and have a translucent pulldown menu show up when you mouse over the image that offers to add a random string of digits to your contacts as an international phone number. Note that it doesn’t let you copy the string; it’s so insistent that it’s a phone number that it only offers options to use it as one.

It’s one thing to be able to open an image in Preview.app and deliberately choose the OCR text-selection mode (which is quite good, even at vertical Chinese/Japanese), but having it turned on system-wide for all images is intrusive and dangerous bullshit. I don’t want every image processed by a text-scanning system that has opaque privacy rules and no sense. And of course interpreting random digits in text as phone numbers and converting them to clickable links is also dumb as fuck; remember when a bunch of DNA researchers discovered their data was corrupted by Excel randomly turning genes into dates?

(Settings -> General -> Language & Region -> Live Text -> offdammit)

(I didn’t even specify an Apple product, just “silver-colored laptop”; training data, whatcha gonna do? I did have to add a table and ask for a big downward swing of the axe, but the flames were free, thanks to a generous interpretation of the term “fire axe”)

Vibe Me Wild

(classical reference)

There is an executive push for every employee to incorporate generative AI into their daily workflow. I’m sure you can guess how I feel about that, but the problem is that they’re checking.

We have licenses for everyone to use specific approved tools, Which I Will Not Name, and VPs can see how many people have signed into the app with SSO, and at least get a high-level overview of how much they’ve been using it.

So I need to get my Vibe on. The problem is, it’s just not safe to run tool-enabled and agentic genai on my work laptop (especially while connected to the VPN), because I have Production access. The moment I check the boxes that enable running commands and connecting to APIs, I’d be exposing paying customers to unacceptable risks, even though there are passwords and passphrases and Yubikeys to slow things down. All it needs to do is vibe its way into my dotfiles without being noticed. I don’t even want the damn thing to read internal wiki pages, because many of them include runbooks and troubleshooting commands. And of course there’s incidents like this, in which an OpenAI “agent” is subverted to exfiltrate your email.

But I need to show that I’ve used the damn app to produce code.

So I wrote up a detailed design document for a standalone Python script that implements my five-star deathmatch image-ranking system. And before handing it over to the work app, I fed it to offline LLMs.

First up, seed-oss-36b, which has been giving me good creative results and tagging: it ‘thought’ for 20+ minutes, then generated a full project that ignored about half of my requirements, including the one about persisting the state of the rankings to disk. I didn’t even try to run it.

Next, gpt-oss-20b, which ‘thought’ for 30 seconds before spitting out a complete, self-contained Python program that almost worked. When I told it that the /images route and the /api/images route worked, but that the main / route displayed nothing and none of the key bindings worked, it ‘thought’ for 25 seconds, realized that it had written syntactically invalid Javascript for the key bindings (multiple case statements on one line), and corrected the code.

At this point, I had basic functionality, but found three flaws in testing. I listed them out, and after 2 minutes of ‘thought’, it corrected them. Sadly, it also deleted the line import os from the code, breaking it.

I told it to fix that, add a new keybind to reset the display filtering, and fix a newly-discovered bug in the image-zoom code that prevented scrolling. A minute of ‘thought’ and it took care of those issues, but deleted the Flask import lines this time.

A mere 7.5 seconds of ‘thought’ convinced it to add that back, and then I had a fully-functional 413-line self-contained app that could let me quickly rank a directory full of image files and persist the rankings to disk.

All in all, ~20 minutes of me time to write the design doc, 4 minutes of ‘thinking’ time, plus ~4 minutes each pass to generate the script (I’m getting ~5 tokens/sec output, which types out the code at roughly twice human speed), plus ~20 minutes of me time for source control, testing, and debugging. Both models used about 18KB of context to accomplish their task, which means that additional enhancement requests could easily cause it to overflow the context and start losing track of earlier parts of the iterative process, with potential loss of functionality or change of behavior.

With tested results, I’m now willing to present the revised design doc to the licensed tool and let it try to write similar code. While I’m not connected to the VPN…

(I suppose HR would take offense if I pointed out that the Vibe in Vibe Coding should be replaced with a more intrusive sex toy…)

With apologies to The Beatles…

I once had a vibe
    Or should I say
It once vibed me
    It wrote all my code
Then gave away
    API keys

It asked for my keys
    and it said they’d be safe in the vault
Then I looked around
    and I found them shared on ServerFault

I called for support, waited online
    wasting my time
I talked myself blue, tech support said
    “I’m off to bed”

He told me his shift had just ended
    and started to laugh
I emptied my wallet
    and crawled off to sob in the bath

And when I came out, my app was gone
    my credit blown
So I set a fire
    at their HQ
and watched them burn

I’d rather find a second waifu than write about evil right now…

(I found her in a bulk-wildcard-testing run where I set it to use the default sampler/scheduler, CFG=3.5, Steps=37, which took 1.5 minutes/image; cranking the steps up to 60 and using Heun++2/Beta improved quality at the cost of taking 6.75 minutes; refining it at 40% and upscaling it by 2.25x with 4xNomosUniDAT_bokeh_jpg eliminated the artifacts and significantly improved detail, but took 25.75 minutes. Worth it?)

(trying to do a 4x upscale ran out of memory after 45 minutes…)

Fiverr lays off 30% of employees…

…to focus on AI. Meanwhile, Fiverr freelancers are making a bundle cleaning up after AI.

🎶 🎶 It’s the cirrrrrrcle of liiiiiiife 🎶 🎶

Wow, Your Forma was worse than I thought…

Some reviewers made a fuss about the chemistry between the two leads being due to their voice actors being married in real life. Not any more.

Maybe she wasn’t willing to do the Sylphy voice in bed… 😁

(official art from the light novels, where she’s a busty little chibi; I’ve already used up all the fan-art, and I didn’t have any luck with the LoRAs)

Now that’s a knife dock!

OWC is selling a Thunderbolt 5 dock with 3x USB-A 3.2 ports, 1x USB-C 3.2 port, 4x Thunderbolt 5 ports (1 used for upstream), 1x 2.5Gb Ethernet port, and 2x 10Gb Ethernet ports with link aggregation support. For those who need their Mac Mini to have multiple 8K monitors and serious NAS bandwidth.

It’d be kind of wasted on me, so I’m thinking I’ll settle for the CalDigit Thunderbolt 4 dock: 4x Thunderbolt 4 (1 upstream), 4x USB-A 3.2.

Out of context

“Reasoning” models tend to do better at prompts like “give me a list of N unique action poses”, except when they dive up their own assholes and start generating reams of text about the so-called thinking process they’re following faking. I’ve had them spend 10+ minutes filling up all available context with busywork, without ever returning a single result. If they do return something, the list is often non-unique or cut off well before N.

But the context is the real problem, because if you only got N/2 results, you’d like to coax it to finish the job, but there’s a good chance that the “thinking” has consumed most of the context, and most models do not support purging context in the middle of generating a response.

It would also be useful to copy the setup of a known working chat and generate a completely different kind of list. I definitely don’t need the context of a search for poses and lighting setups when I ask for festive holiday lingerie ideas.

You can’t just reload the model with larger context and continue the chat. You can’t fork the chat into a new one; that preserves the context, too. What you want is to wipe all answers and context and preserve the settings and your prompts, so you don’t have to set up the chat again. In LM Studio, this functionality doesn’t seem to exist.

So I wrote a quick Perl script to load the JSON conversation history and remove everything but your prompts.

#!/usr/bin/env perl
use 5.020;
use strict;
use warnings;
use JSON;

open(my $In, $ARGV[0]);
my $json = join(' ', <$In>);
close($In);
my $conv = decode_json($json);

# remove token count
$conv->{tokenCount} = 0;
# rename
$conv->{name} = "SCRUBBED " . defined $conv->{name} ? $conv->{name} : "";
# remove all messages not from user.
my $out = [];
foreach my $msg (@{$conv->{messages}}) {
    push(@$out, $msg) if $msg->{versions}->[0]->{role} eq "user";
}
$conv->{messages} = $out;
print encode_json($conv);

This dumps a fresh JSON conversation file to STDOUT, which can be given any filename and dropped into ~/.lmstudio/conversations.

Why Perl? Because I could write it faster than I could explain the specs to a coding LLM, and I didn’t need multiple debugging passes.

(BTW, 32K context seems to be a comfortable margin for a 100-element list; 16K can be a bit tight)

Digits and Shadows

It’s kind of amazing how the “state of the art” keeps advancing in AI without fixing any of the known problems. LLMs hallucinate because they must, and image-diffusion engines can’t count fingers or recognize anatomical impossibilities because they don’t use anatomy. All the alleged coherence you hear about in AI output is either sheer accident or aggressive post-processing.

Let’s examine a few types of Qwen Image failures:

“Thumbs, how do they work?”

Trying to put your own hand into a position seen in a generated image can be awkward, painful, or simply impossible if it’s on the wrong side.

more...

(we now resume our regularly scheduled randomblogging, until the next slaughter of innocents by Left-wing domestic terrorists)

Boxxo Or Bust 2, episode 11

This is really not a good week for the show to go grimdark, revealing Our Shoe Fetishist’s bloody, revenge-fueled path. I’m just not going to watch it right now.

Here’s something wholesome instead:

Warner Bros wants to be the Broadcom of streaming

CEO insists HBO Max should be priced like diamonds. Six customers are so much easier to support than six million.

Not a fan

I bought a new unmanaged 10Gb switch, because my existing one only had 2 10Gb ports with the rest being 2.5Gb. Three years ago, the price of a switch with 8 10Gb RJ45 ports was a bit too spicy, but now it’s just annoying.

What really hurts is when a desktop switch has a fan loud enough to be heard from the next room. And I need it to be in the same room. For comparison, I barely notice the Synology NAS that’s on the same desk, even when it’s doing a RAID scrub. For the sake of my sanity and productivity, I guess I’ll be building the switch a padded crib…

New Monitor(s)

The new monitor is an “ASUS ZenScreen Duo OLED 360° Foldable Portable Monitor (MQ149CD)”, which is quite a mouthful, but the TL/DR is that it’s two 1920x1200 HDR panels that can be used separately or together, portrait or landscape. To use it as two separate monitors, you need to use two cables, but despite the one-inch gap between the panels, you can still get good use out of it as a single 1920x2400 display.

I’m using the 1/4”-by-20 tripod screw to mount it higher than the kickstand, and despite all claims of portability, yeah, you have to plug it in to drive it properly (1 of the 3 USB-C ports is power-only). Right now I’ve just got it hooked up with HDMI, because my current dock was low on free USB-C ports, but once I get things rearranged I’ll separate them, because Lightroom’s window layout is a bit inflexible, and the one-inch gap is bad for photo editing. Better to use the application’s “secondary display” support.

This does make me want a vertical 20+ inch 4K HDR portable monitor, though. I just don’t want it to be from a company I’ve never heard of before, which is what’s all over Amazon.

“Shared memory and fancy NPUs are no match for a good graphics card at your side, kid.”

While I’m discussing new toys, I’ll mention that I also picked up a refurbished M4 Pro Mac Mini, which in theory has much faster “AI” performance than my M2 MacBook Air. With 64 GB of RAM, it can run mid-sized offline LLMs at a decent pace, and even do quick image captioning if I downscale raw images to a sensible resolution.

What it can’t do is more than very basic image generation. It’s not the memory, it’s the cores: not enough of them, and not fast enough. This matches what I’ve heard about the insanely-pricy Mac Studio and the merely-way-overpriced systems built on the AMD Ryzen AI Max+ 395. Their performance comparisons are based entirely on running models that simply don’t fit into a consumer graphics card’s VRAM. “Our toaster is much faster at running this 48 GB model than a 24 GB Nvidia 4090!”

The Mini is a huge speedup for Lightroom and Photoshop, though, which made it worthwhile for me. Any “AI” playability is a bonus.

(I have never seen a human being hold a camera this way)

Related, “Dear Draw Things programmers”

Come back when you discover the concept of legibility. Dipshits.

Accidental Disney Princess

I was testing a few more LLM-enhanced dynamic prompts, and one in particular stood out: a series of very consistent images of a fresh-faced young gal who could make Bambi’s woodland friends sit up and beg.

4K resolution, crisp and highly detailed, create an illustration that exudes a high-budget aesthetic. Depict an average height, lovely ethnic Egyptian woman who is 18 years old with a petite figure. She has deep blue eyes, lobed ears, a straight nose, a wide chin, an U-shaped jaw, dimpled cheeks, and a flat forehead. Her heart-shaped face highlights her gentle features. Her skin is healthy and alabaster white, with festive holiday makeup that complements her almond-shaped eyes and full lips. Her hair is steel gray, styled in a cybernetic pixie cut with metallic edges and glowing circuit-like patterns. The woman has a happy expression as she stands with one arm raised, bathed in luminous waves of light. The background features a lush jungle filled with singing flora, glittering with raindrops under the radiant colors of a rainbow sun. Subtle dawn light filters through, creating an aerial perspective with open framing and pastel tones that evoke a gentle awakening. The composition highlights her graceful form against the vibrant greenery, capturing a dreamlike atmosphere.

Sadly, the prompt that worked so well in Qwen Image was… less successful with other models that don’t have an embedded LLM to parse complete sentences sensibly, and failed completely with a few that couldn’t handle the combination of the Beta scheduler and the Heun++ 2 sampler (I may redo those with their defaults).

It’s not that the others all failed, they just didn’t produce consistent characters who looked like my princess and/or lived in her enchanted forest.

more...

(…that’s not another season of Frieren or Apothecary Diaries)

The World’s Strongest Rearguard anime

I consider this to be one of the best cheat-power isekai light novel series. Our Ass-Guardian Hero reincarnates into a fantasy world that basically forces all the newcomers into dungeon crawls without any real choice or training, but while he’s being railroaded by a hot chick with big boobs, he accidentally chooses a rare class. Everyone else is trying to be things like “valkyrie”, “soldier”, “forager”, “tennis coach”, etc, but hearing that parties are always looking for someone to take the rearguard position, he just writes down “rearguard”.

And he gains the ability to buff any party member who is in front of him. It should come as no surprise that he ends up with a party of gorgeous gals, but it might be surprising that he is a complete gentleman who doesn’t so much as stare at the asses he’s guarding. He’s not oblivious to their (mostly-underage) charms, but he never, ever goes there.

What makes it work is that despite him being potentially and eventually overpowered, he and his party have to work for every fight, and their trust and teamwork keep them alive just as much as their skills. His power gives them an edge, but his decency and kindness have just as much of an impact on the world.

Add to that that every not-quite-haremette has a distinct personality and goal, and you have something that I’m begging the animators not to screw up.

The buried lede in this article is that book nine is finally coming out after a four-year wait. Book 8 wrapped up a major story arc, but left a lot of character arcs unresolved, so woo-hoo!

(Our Hero is in the background, with His Formerly Abusive Former Manager on the left, Our Terrifying Teen Swordsgal on the right, and Our Best Girl Rescue Lizardgal in the middle; and the butterflies are seriously bad news, with their presence suggesting two cours)

Just-Outta-Time Logistics

I bought an HDR monitor on Amazon, which was supposed to arrive Monday. At 4pm, they announced that for unknown reasons it never left an Ohio warehouse, and it would instead arrive Tuesday. Three hours later, that message disappeared, and the order status reverted to arriving-Monday-but-not-shipped-yet. It stayed like that until after 10pm, when it went to “delayed, not shipped” again, with no further explanation.

In the wee hours of the morning, Amazon’s entire order-history system apparently went offline, and all status was unavailable.

As of 8:30 AM, it is allegedly across town, waiting to get put on a truck. Unless they erase that status update.

Unrelated,

FedEx grudgingly coughed up an approximate delivery time for Tuesday’s signature-required package (not from Amazon), which is good, since I have to pick up my niece after school, and I’d hate to have to go chase down the package at their nearest office after the truck returned to base.

Amusingly, when I asked for details, they offered two confirmation methods: SMS to my (correct!) cellphone number, or email to a masked address that I didn’t immediately recognize. It would have reached me anyway, since I own the domain, but it took me a moment to figure out that they’d decided that the email address associated with my physical address was the one I’d created specifically for B&H Photo.

Because they shipped me a CompactFlash Express card (fastest storage for my 4k-video-capable Sony A7SIII) via FedEx a while back.

Mistral Small is gaaaaaaaay…

Given the following picture of Yuzuha Hongo:

And the relatively-capable vision-enabled LLM mistral-small-3.2, we get the following caption and tag set that contain a rather serious error:

A young woman with long brown hair, straight hair style, and a slim body type is standing in a bathroom. She has fair skin, small breasts, and is wearing a pink floral bikini top with thin straps. She has a slight smile on her face and is looking directly at the camera. The bathroom features white tiles, a shower head, and a towel rack with towels hanging on it. Water droplets are visible on her skin.

Tags: young woman, long brown hair, straight hair style, slim body type, fair skin, small breasts, pink floral bikini top, thin straps, slight smile, direct gaze, white tiles, shower head, towel rack, hanging towels, water droplets

(ran it a few more times, and got completely different results every time; sometimes it correctly handled her boobs (coughcough), other times it decided her nipples were visible, several times it hallucinated a towel rack with towels, occasionally it remembered to mention “Asian facial features”, etc, etc; honestly, I trust the output so little that I’m tempted to have my script loop over the same picture N times and keep only the tags that show up at least 50% of the time; I’m also going to borrow the styling of the Qwen prompt-enhancer below)

Enhanced dynamic prompting

With my new still-in-progress dynamic prompt collection, I’m turning this:

__style/photo__ of a __var/prettygal__ with __skin/_normal__ and __hair/_normal__, with a __mood/_normal__ expression. __pose/sexy__. Her location is __place/_normal__. __scene__.

into prompts like this:

4k, crisp, high budget, highly detailed, intricate, ultra textured. Create a photograph of a average height, beautiful, ethnic American, adolescent woman with supple figure, Steel gray eyes, softly-curled lobe ears, Angular Nose, V-Shaped Chin, Strong Jaw, Angular Cheeks, Angular Forehead, Triangle face shape, Natural “no-makeup” makeup with healthy Sun-Kissed-Tan skin and Natural Gray hair, Low twist with braids, creating texture and detail, with a warm expression. Standing with weight on one leg, other foot planted flat, arm crossed over chest, wrist aligned, chin up. Her location is Grand Baie, Mauritius. Rainy glow, close framing, layered reflection, muted palette, melancholic ambiance.

Then using LM Studio’s Python API to pass it through qwen2.5-7b-instruct, with this (taken from the demo code in their repo):

And getting enhanced prompts like this:

4K resolution, crisp and highly detailed, with a high budget aesthetic that emphasizes intricate texture. Create a photograph of an average-height, beautiful ethnic American adolescent woman with a supple figure. Her steel gray eyes sparkle softly, and her ears are gently curved. She has an angular nose, a V-shaped chin, a strong jaw, angular cheeks, and an angular forehead, forming a triangle face shape. Her natural “no-makeup” makeup highlights healthy sun-kissed tan skin, while her naturally gray hair is styled in a low twist with braids that add texture and detail. She stands with weight on one leg, the other foot planted firmly, and her arm is crossed over her chest with wrists aligned. Her chin is lifted slightly, giving a warm expression. The setting is Grand Baie, Mauritius. The photo captures a rainy glow, using close framing to create layered reflections in the background, maintaining a muted palette that evokes a melancholic ambiance.

Now, let’s see what sort of picture that produces…

more...

This is Alibaba’s offline text-to-image model, and it has excellent prompt comprehension (LLM-based), high image quality, a higher-than-usual base resolution, and is capable of some naughtypics. It is also quite large (20GB+) and slow, especially when you crank up the parameters to render small text. This image took five minutes to render on my high-end gaming PC:

(the main title always came out correct; the subtitle had about a 50% chance that all the letters would be intact)

Okay, I can work with this (2-minute render):

It doesn’t seem to have some of the training issues I ran into with Omnigen2, but while it’s not fully censored, it’s also not a true NSFW model, and has a tendency to cover the naughty bits unless you specify your request precisely. Checking Civitai, there are people working on LoRAs to provide explicit content, but it takes a while to figure out the quirks of new models. Y’know, if that’s what you’re into…

more...

Seasonal Gushing gals

As we wave farewell to Leoparde’s summer sizzler…

…we look forward to the trick-and-treat cast party:

Offline image generation: Omnigen2

A fairly new model, Omnigen2 is supported by SwarmUI for image generation and editing, so I don’t have to change my workflow to play with it. It has a full LLM embedded, so it’s better at parsing prompts than the usual Stable Diffusion models, but still has a learning curve. Compared to ChatGPT, you’ll spend a lot more time writing a lengthy, detailed prompt that produces the results you want, but won’t get cockblocked by randomly-changing secret policies that refuse to generate your picture at all.

It is definitely capable of producing naughtypics, but is not overtrained to the point that it just goes there at random.

So, how does it handle the same prompt that produced my latest fake isekai book cover?

more...