Every morning at 4 a.m. Claude draws one idea with three image models, and it has picked the same one 41 times out of 41
A scheduled Claude job borrows my laptop's GPU before I wake and draws one idea with oneObsession, Pony V6 XL and Qwen-Image 2.1, plus a short video clip. In 41 ideas its judge has never once named anything but the Qwen picture the best. Six of those mornings are below, shuffled, so you can judge them blind.
Every morning at four, a scheduled Claude job checks that nobody is using my laptop’s GPU, then draws one idea with three image models and turns the best still into a five-second clip. It looks at every picture, writes a sentence about each one, and names a best of the idea. In 41 ideas since 24 September, the best has been the Qwen-Image 2.1 picture every time.
That is less a verdict on the models than on the judge. Six of the mornings are below, with a button that shuffles them and hides the names, so you can pick before you know.
Where it came from
On 20 September I told Claude what I actually wanted from all the agents running on my machine:
What I really want is Aaron candy. As in things that give me joy and signal.
The first run went out the next morning: one model, Flux.2 Klein 9B, three posters for a racing game I’d made that never had cover art. By 24 September the job had its current shape. A Claude Sonnet 5.5 agent picks one new idea, from one of my characters, one of my projects or something in my notes, and makes it four ways:
- oneObsession, an Illustrious checkpoint, prompted in comma-separated tags;
- Pony V6 XL, also tags, behind its
score_9, score_8_uppreamble; - Qwen-Image 2.1, prompted in plain sentences, the one that can letter a sign;
- MiniMax-H3, a five-second clip with sound, animated from the best still.
The clips stay off this site: MiniMax-H3’s licence excludes the United States, for the model and its outputs. Everything else here is the job’s own output, resized for the web.
Every generation gets a line in a log before the next one starts: the time, the model, one sentence about what the picture actually shows, the measured seconds, and kept or dropped with a reason. That log is where every number in this post comes from.
Judge them yourself
These are six of the eleven mornings since 1 October, each model’s better picture of its two by the job’s own notes. Every row is one brief. Press Judge blind, pick one per row, then Reveal: the job’s pick is marked.
The same idea, in three dialects
The difference starts in the prompts. Here is what the agent wrote on 10 October for the Halloween row, oneObsession first, then Pony, then Qwen:
1boy, solo, male focus, tall, gaunt, pale skin, black hair, (iron faceplate covering
eyes:1.4), (huge grin:1.3), white victorian shirt, tattered red cloak, holding candy bowl,
standing, porch, halloween, jack-o'-lantern, small children in ghost costumes, night,
(quasarcake:0.65), limited color
score_9, score_8_up, score_7_up, score_6_up, source_anime, rating_safe, 1boy, tall gaunt
man, pale skin, black hair, (iron plate covering eyes:1.3), (huge wide grin:1.3), white
victorian shirt, long tattered red cloak, holding candy bowl, porch at night,
jack-o'-lantern, halloween, small children in ghost costumes, painterly
A painterly Halloween poster in a desaturated grey-blue night with the red cloak as the only
saturated colour, plus the warm orange glow of jack-o'-lanterns. A tall gaunt pale man stands
on a wooden porch, wearing a smooth featureless iron plate over his eyes with no eye holes,
and an enormous gentle grin far wider than a human mouth below it. [...] He holds out a
wooden bowl heaped with wrapped candy toward three tiny children in bedsheet ghost costumes
who cling together at the bottom of the steps, frozen with fear, one of them reaching out a
trembling hand. [...] At the top of the poster, a cream banner with hand-lettered text
reading: PLEASE, TAKE ONE.
A tag list can say there are children in ghost costumes. It can’t say they are scared of him, or that one of them is reaching out anyway, and that relationship was the whole idea. Neither tag model drew a frightened child. oneObsession put a pumpkin on his head in its first try.
What the log says
| Model | Kept | Dropped |
|---|---|---|
| Qwen-Image 2.1 | 77 | 2 |
| oneObsession | 46 | 8 |
| Pony V6 XL | 22 | 21 |
The seconds are close. The median render was 21.3 s for Pony, 26.7 s for oneObsession and 28.8 s for Qwen; the clip takes about 81 s. A whole morning is seven generations and four minutes of GPU, and the agent’s tokens cost $1.75 to $2.87 a morning at API list prices since 1 October. On 4 and 12 October it found the GPU already busy and stood down.
On 30 September I told it my GPU was free and to keep going until I said stop. It ran nine more rounds, 27 ideas in about two hours, before I typed “stop for now”. That afternoon accounts for most of the 41.
The judge grades against the brief, and the brief is written in sentences. Read the
dropped lines and the reasons are almost never about how a picture looks. Pony’s Blackwater
poster on 25 September was “the most cinematic and genuinely unsettling of the four despite
the anglerfish being dropped from the render”. Its Playback Lens picture was “beautiful but
the lens, the scroll and the idea are lost”, and was dropped. Ichigo’s autumn morning lost its
Pony picture because a yellow snail tag turned into “two empty yellow rowboats each carrying
a small animal-like head, no girl anywhere”.
So the 41 doesn’t say Qwen draws the best pictures. It says Qwen draws the picture that was asked for, including the words on the sign, and that an agent holding the request will pick it every time. The job still makes all three every morning, because its instructions say to. Going by its own notes, the other two are where the surprises come from: the pictures nobody asked for, which it keeps, describes kindly, and never chooses.