| Can a terrain model make real mountains? | Rank-8 LoRA on a diffusion coarse stage | Blurry ocean to ridges and fjords in 400 to 600 steps |
| Is the terrain Earth-like? | Pretrained cascaded latent diffusion | 10 of 12 statistics match; river exponent 0.21 vs 0.38 |
| Can fine-tuning fix the rivers? | Base-stage LoRAs; retrained encoder | Loss fell while realism got about 5× worse; closed |
| Can I steer it without training? | Best-of-N and SMC with a reward gate | 0 of 17 candidate rewards admissible |
| Train it from scratch? | Pixel-space EDM U-Net, 18M and 50M; SPADE conditioning | No collapse; SPADE 2.0× stronger; gates missed |
| Learn how caves are shaped? | Sequence diffusion, residual EDM, score model, GRU, MLP operator | Hybrid kept; simpler models won where diffusion lost |
| Turn concept art into game assets? | Hunyuan3D-2.1, native PBR painter, RigAnything, SAM 3, Blender | 12 inputs to rigged, animated GLBs, fully local on AMD |
| Can an image model paint in my art style? | Style LoRAs on FLUX.1-dev, SD 3.5 and FLUX.2 Klein (ai-toolkit) | 3,250-step FLUX LoRA on 300 captioned images, used for concept art |
| Can the style carry into video? | Style LoRA on Wan 2.2 Animate 14B, both noise experts | Trained 1,000 steps and applied to a talking-head clip |
| Can a small model hold a personality? | Rank-64 LoRAs on Qwen3.5-9B and Gemma 4 E4B from a synthetic data pipeline | About 350 filtered examples per character; voice transfers, no scored eval yet |
| Teach realism with a simulator? | Stream-power erosion solver as a synthetic teacher | Calibration failed on all 43 settings; no-go |
| Feed a neural world model? | Headless procedural driving, 50 episodes × 1,000 steps | Dataset built in DIAMOND format; model not yet trained |
| Can a tiny model sound like my game server? | 3.8M-parameter GPT on 13.6M tokens of server chat, ONNX in the browser | Runs client-side in a browser sandbox |
| Serve a 122B model on a 128 GB APU? | MoE with multi-token-prediction vs DFlash speculative decoding | MTP 25–29 tok/s; DFlash 8–13, slower than no draft |
| Trust a vendor's quantized model? | 4-bit 27B, HumanEval audit | Speed held; quality 154/164 vs claimed 158/164; not adopted |
| Can an LLM play my modpack? | Local 9B planner over a deterministic game controller | Passed 29 of 30 screens, approved 1 of 7 bad proposals; kept out of autonomous play |
| Find anything across agent notes? | Multimodal embeddings (Jina v5) behind an MCP server | Text index in daily use by the agents |