Eric Yu (@ericfish) header
Eric Yu (@ericfish) profile picture

Co-Founder @MathWallet

Singapore
Joined April 2007
Before turning an image into a 3D asset for Wonder Town, I tested the image itself. Gemini 3 Pro Image via OpenRouter vs Qwen Image 2.1 and MiniMax H3 Image in ComfyUI on my LAN. Nov 18 test: 4 characters and 2 landmarks. The cloud baseline uses existing assets; local runs used our pipeline prompts. I also tried style references and stronger proportion wording. MiniMax was closest to our vinyl-toy style in this sample. Children, robots and monsters looked usable; the adult villager came out tall and thin until I supplied a style reference. Qwen improved the villager's proportions with stronger wording—but lost his glasses. Its yeti kept a furry texture. MiniMax's landmarks had cleaner, simpler shapes. That looks promising for image-to-3D, but I haven't tested the conversion yet. For now, concept images stay on OpenRouter. Cost was about $0.14/image in this workflow; local runs still have hardware and compute costs. One seed per image, prompts originally tuned for Gemini, no 3D conversion test. A workflow check, not a general model ranking.
3
3
256
I compared Hunyuan3D-2 on my LAN with Tripo for Wonder Isles. Same concept image for each asset; 4 characters + 2 landmarks; one run per asset. My choice: • Landmarks → Hunyuan. The book tower was close enough; the lighthouse still lost rear details. Saves 30 Tripo credits per generation. • Town NPCs → Hunyuan + local Blender rigging, with Tripo rigging as fallback. • Playable characters → keep Tripo. The camera spends too much time behind them to ignore back-view quality. Mia's backpack and the librarian's glasses made the trade-offs obvious. Front views alone can hide a lot. Hunyuan generation + local rigging uses 0 additional Tripo credits; compute costs still exist. Local rigging reuses an existing rig and animations, so body shape matters. Small workflow test, not a general ranking: exact Hunyuan variant unconfirmed; mesh budgets differ; rigged outputs only checked in a separate viewer, not in-game. Images: Tripo top / Hunyuan bottom. Tested Nov 18.
3
1
82
Added seasons, weather, and a time system to Wonder Isles today. 🌦️ What might once have taken me a month now takes a few prompts. Two skills still matter: 1. Defining exactly what to build. 2. Understanding the limits of frontier models.
36
Built this little town with Fable 5.2 today. Planning to turn Wonder Isles into a birthday gift for my daughter. 🏝️ Now for the fun part: what should she be able to do here? Treasure hunts, make friends with the villagers, decorate a home? What would you add?
1
2
158
Execution keeps getting cheaper. The ability to see the cathedral keeps getting more valuable. Same wall. What decides your fate isn't how fast you lay bricks — it's what you see when you look up. Building a wall, or a cathedral? 👇 #AI #FutureOfWork
4
22
So — which mason are you? AI doesn't erase an industry. It erases a way of understanding work. • See only the brick → in danger • See only the money → still in danger • See the whole cathedral → AI becomes your strongest crew
1
4
36
Layer 3: hardest to replace — the cathedral designer. Same wall, but he holds the whole cathedral in his mind. He doesn't perform an action; he answers a question: • Define the problem • Give it meaning • Own the judgment where there is no right answer
1
6
Layer 2: AI even replaces the "breadwinner." The second mason's whole idea of work is "a thing that pays." Bricks, boxes, words — just different ways to trade hours for money. But AI is rewriting what work is worth. Treat a job as only a paycheck, and you vanish with it.
1
10
Layer 1: AI replaces the bricklayer first. The first mason sees only the brick in his hand. To him work is a loop — haul, spread, align, repeat. Anything that breaks into clear steps with a right answer gets automated first. Execution itself is becoming free.
2
2
30