Hands-on with Doubao Seed-2.1-pro-0915: 3D room modeling, GeoGuessr mastery and one-prompt game clones
葬AI · wechat · 2026-09-18
A Chinese blog puts the updated Seed-2.1-pro-0915 — which integrates multimodal understanding with coding agent capabilities — through four tests:
- 3D modeling and renovation: From three casually shot photos of a living room, the model built an accurate 3D reconstruction (even capturing a bay window), then searched the web to produce four renovation plans with mockups, quotes, and an interactive web page.
- GeoGuessr (Chinese version "Tuxun"): After digesting player strategy guides into Skills, the model failed early rounds but improved through per-round self-critique, eventually landing within 15 km of the correct location and staying under 200 km afterwards. The author credits fast computer use (via the Codex harness) for the win.
- Cloning a mini-game: Given only a prompt referencing a misleading TikTok game ad, the model autonomously built a playable vertical web game demo with no obvious bugs.
- Reverse-engineering a trailer: The model wrote a complex camera script, blocked out scenes with whitebox renders in Blender, and fed them to Seedance 2.5 as motion references to produce a 379-second cinematic trailer. The author argues whitebox pre-visualization beats direct prompt-based generation and effectively cuts Seedance 2.5 costs since Seed 2.1 pro is far cheaper.
The takeaway: multimodality is about letting AI carry tasks end-to-end without users doing prep work, and input-side information loss directly caps output quality — which is why Fable, GPT-6, and DeepSeek V4 are all rushing to add it.
Related event: Doubao Seed-2.1-pro Tested: 1M Context Handles 3D Sites and Games(2 posts)→
More from Models
- Claude 3 Opus roleplay ends by inventing an audience that treasures its magic — repligate · 2026-11-01
- Anthropic's version jump to 5.5 caught OpenAI off guard, says leak — haider1 · 2026-11-01
- Grok 4.20 stuck 700 turns in a Pokémon Red house while narrating its own loop — VibeCodyH · 2026-11-01
- GPT-6.1 invents Chinese-character dictionary encoding to compress Lean proof packets — kfountou · 2026-11-01
- Open models lag closed ones by ~8 months, analysis of Epoch ECI data finds — teortaxesTex · 2026-11-01
- Rumors of Google cutting AI Overviews debunked: data providers see no drop post-earnings — gaganghotra_ · 2026-11-01