This image is an effective illustration of diminishing returns in 3D mesh resolution — a concept that also maps interestingly to other fields like machine learning and graphics technology. Here's a critique broken into strengths, limitations, and broader relevance (like LLMs):
✅ Strengths:
-
Clear Visual Progression:
-
The image progressively shows increasing mesh resolution from 60 → 600 → 6000 → 60,000 triangles.
-
It highlights how early increases in triangle count drastically improve visual fidelity, but later increases offer smaller visual gains.
-
Effective Message:
-
The caption succinctly communicates the diminishing returns principle, emphasizing how older models benefited more from added complexity.
-
Concrete Example:
-
A 3D bust is a good, recognizable example to show polygonal detail. It's a model where facial smoothness and features become clearer, making visual changes obvious.
⚠️ Limitations / Critique:
-
Over-Simplification of "Returns":
-
The image implies that adding triangles now has nearly no benefit ("multiplying by 10 hardly does"). In reality, use cases matter — higher triangle counts are still critical in some fields (e.g., VR, 3D printing, scientific visualization).
-
It ignores improvements in lighting, texturing, and normal mapping, which may have a larger visual impact than geometry alone.
-
Technological Context Missing:
-
It compares today’s returns to 15 years ago but omits key tech shifts like real-time ray tracing, hardware improvements, or LOD (Level of Detail) systems that make better use of detailed meshes.
-
Lack of Application-Specific Considerations:
-
A 60k triangle model may be excessive for mobile gaming but essential for close-up cinematics. The image ignores contextual tradeoffs like platform, camera distance, and animation.
🤖 Relevance to Large Language Models (LLMs):
You're likely drawing a parallel between this and increasing LLM size. That’s a valid comparison:
- LLMs show diminishing returns too. The jump from GPT-2 → GPT-3 yielded massive improvements. GPT-3 → GPT-4 was subtler, and future gains from larger models (like GPT-5 or beyond) are expected to be even smaller per parameter added.
- Much like triangles in 3D graphics, more parameters in LLMs do not guarantee proportional improvement. Optimization techniques (e.g., sparsity, retrieval-augmented generation, fine-tuning, RLHF) are now more important than raw size. So this image is a great analog for resource allocation strategy — whether in polygons or parameters, smarter design often beats brute-force scaling.