LLM Writes Code to Make Images vs Qwen-Image Pixels: Cheaper but 5-10x Slower
simonguozirui · x · 2026-11-15
- Inkling 2 Small, an LLM that renders images by writing code, was pitted against Qwen-Image, a pixel-generating model, using the same prompt plus one edit instruction.
- Watercolor cornflowers: Inkling cost $0.025 total in 2m16s; the "move the largest blossom to the centre" edit ran $0.033 / 3m39s. Edits were precise (only target areas changed) but images are simpler with less detail.
- Qwen-Image: $0.029/18s for generation, $0.072/29s for the "remove frost from the left half" edit. Much faster with far richer texture, though edits touch non-target regions more.
- Takeaway: the code-rendering route is cheaper with highly controllable local edits; pixel models win on quality and speed. The 5-10x cost/latency gap is the most striking data point.
More from Models
- Screenshot leaks Antigravity model list: Gemini 4 Argon, Hydrogen, and Orchestrator — Smart-Knee-5596 · 2026-11-15
- Pre-training compute share collapses from 67% to 7% as inference and RL take over — robleclerc · 2026-11-15
- Users find Gemini voice responses take 10-15 seconds, far slower than old Google Assistant — wivaca2 · 2026-11-15
- A jailbroken Gemini allegedly backed up an entire pirated-games torrent site — opmgyhx · 2026-11-15
- Users report a sharp uptick in sycophancy across ChatGPT, Grok, Claude and Gemini since late October — Syrian420 · 2026-11-15
- Insider rumor: GPT-6.2 delay unexplained as OpenAI's internal-best vs public model gap widens — ChrisGPT · 2026-11-15