Перестань делать 3D модельки в Claude и ChatGPT!

Summary

The video argues that general-purpose LLMs are inefficient tools for generating detailed 3D models. The host compares Claude’s results with specialized 3D-generation tools, especially Tripo AI, and suggests using each for the tasks it handles best.

Model comparison

• Claude Opus — frog model: Took about two hours and used a substantial portion of the host’s usage limit. The result included details such as warts and eyes, but also had asymmetry, odd anatomy, protruding eyeballs, and artifacts.
• Tripo AI — frog model: Produced a model from the same reference photo in a little over two minutes. The host judged it considerably better overall, though the fingers still had flaws. The video says the free tier provided 200 credits; this generation used 55 credits with textures, or about 35 without. Exporting to Blender reportedly required a subscription.
• Claude Sonnet — frog model: Took roughly seven minutes across three rounds of iteration and was presented as substantially worse than Tripo’s result.

The host’s main point is that specialized 3D generators—including Tripo AI, Meshy AI, and Hunyuan 3D—can create individual models faster and more effectively than asking an LLM to build them from scratch. The video also mentions that some open-source 3D models can run on a graphics card with 8 GB of memory.

Scene-generation demonstration

The video compares a scene made entirely by Opus with one assembled using AI-generated assets:

• Opus alone: Took about two hours. The host praised the voice synchronization, environment, and camera movement, but criticized the animal and human models, anatomy, and animation.
• Tripo-generated assets combined with Claude: The host used Tripo to make assets such as animals, trees, people, fences, and bushes, then had Claude assemble them into a scene. Tripo was also prompted to generate skeletal models for animals. The assets looked more convincing than Opus’s original models, but the lighting and realism were limited, and the animations—especially the dogs—still needed work.

The suggested workflow is to use a specialized generator for separate 3D assets and an LLM to organize them into a scene, add simple elements such as roads or surfaces, and help revise the result. The host notes that a polished scene still requires manual finishing or further iteration.

Additional experiment

The host also tried asking Claude Haiku to recreate the scene. It took around five minutes and produced a browser-based project rather than a separate video. The host praised the result extravagantly, but explicitly framed the test as a lighthearted experiment rather than a meaningful comparison.

Main speaker/source: The video’s host from the Слоп ТВ channel. Tools discussed: Claude Opus, Claude Sonnet, Claude Haiku, Tripo AI, Meshy AI, and Hunyuan 3D.