Opus 5.5 sculpted every part of this ship in code, including the dragon, sails, and architecture. I can't say it has bad taste anymore. But it still needs steering, so I tried a new way of prompting. Live site: I asked Opus to score every part of the scene:... the ship, water, landscapes, buildings, and their individual parts. Then I asked it to keep improving each part until it scored at least 8 out of 10, depending on my token and time budget. The key is to let it work as long as possible. The longer it works, the better it gets, until it reaches a point where I need to course correct. For me, Opus 5.5 reaches a level of detail and realism that no other model does.show more

I tried this with fable and got a bunch of threeJS lego men primitives that made absolutely no sense lol

they love minecraft and primitive shapes.

@tomjohndesign has done something interesting by making a primitives editor that the user can use for direct feedback to the model, and bakes sprites from the primitives

For real opus 5.5 is going crazzzzzy!

In my opinion, AI will never one-shot art. While its ability to sculpt 3D models with such fine detail will continue to improve on a single prompt, art itself is creative. I don’t know how AI will ever read your mind and provide you exactly the image you were thinking down to every detail. What is needed is more control over detail where the AI doesn’t completely regenerate the entire image but rather can regenerate aspects of it until it meets what you envisioned. Grok Imagine is heading the right way with that

it will, but it will create a ton of slops too.

@JoeWinton How do I put it without hurting… ehhh… you guys are in an AI-psychosis. While AI can direktly do 3d, blender or tripo it is far below the qualitybar for anything really good. Right now its still in a hobby stage/ basemesh blockout but I would rather spend those tokens on code.

The output is very much early 2000s level graphic quality. But the rate of progress will close that gap in the very near future. I also would argue that tools like blender or Tripp will go away as the LLMs become more capable to achieve results without the use of MCPs or other connectors.

that's a nice boat

all i asked was for a japanese boat.

Opus 5.5 just built a ship with a dragon that looks like it got designed by a drunk architect and a very confused AI.

yeah i could improve that more. this is why we need meshy or tripo3d and a ton of patience for now.

One thing worth testing on the scoring step: reviewer/critic prompts in my agent fleet judge more sensibly at medium reasoning effort than at high or max - cranking it up for the grading pass tends to make the scores worse, not more discerning.

interesting. i only use extra.

this is really amazing

the scoring part is quite important.

Self-grading loops tend to plateau once you hit the model's own taste ceiling, it'll keep reporting 8/10 after real improvement stalls. How many rounds before the score stopped tracking actual quality for you?

Letting it self score and iterate is a smart way to steer without micromanaging.

ooo nice. when I have usage to waste I will try and build a multiplayer sea of thieves but in lego.

Scoring each part to 8 and letting it run is a better loop than one giant prompt. The dragon still needs a human pass.

animals and people are notoriously hard. if only i wasn't out of tokens.

The detail on that dragon is wild. Love that you're treating the steering as a separate layer smart way to harness the power without losing your voice.

Three js God

im getting way more credit than i deserve :)

Is it also in just 3js or blender unreal engine etc. also did u supply asset files ?

all in three.js

Did u hv to design assets scenes separately then clubbed or was it single flow like God, make sea, now ship, mountains....

insane

This is unreal. Is this all in the browser? You haven’t imported any videos or imagery?

all prompting and all code. no library.

Amazing work dude 👏🏼 I need to start doing more of this then.

Don't just stop with the ship. Push it more.

the more ambitious, the longer

I just wonder how many tokens you used

Can't understand what they did with 5.5 but when using Opus 5, if we said Claude can make this it would be a joke.

Shader work on water is the most interesting thing about this tests. If you could make it work further with foman, chore foam and random waves... I have already seen some interesting examples

very beautiful!!

Just curious, what are you telling the AI to do when you reference “keep improving”

wtf

Slop

Rubric plus token budget and it stops plateauing. The trick is the loop, not the model. I will take that over model hype any day.

This is nuts

API or subscription?