Posts

On a Radeon 780M, Qwen-Image-2.1 generated images about 5.5 to 9.3 times faster than Qwen-Image-2512 in this test. It also saved a potion sprite with transparency in the PNG.

Testing Qwen3.8-27B on a Radeon 780M with 24K context. MTP raised decode speed by 67%; a 20K-token input took about six minutes before output began.

A local model pipeline on a Radeon 780M planned, built, tested, and repaired a mobile tower-defense game, exposing why executable browser evidence matters more than model self-review.

Generating article backgrounds with Qwen-Image-2512 on a Radeon 780M, then adding English and Japanese titles for Astro Open Graph images.

The Gemma-4-12B encoder and the diffusion model can't share 14GB of iGPU memory, so ComfyUI runs the encoder on CPU and every new prompt costs 20 minutes before the first frame. Splitting the encode into stable-diffusion.cpp fixes that, once you find the model stage it silently skips.

Testing a 20B-class open-weight image model on a Radeon 780M iGPU with no discrete GPU. Qwen-Image-2.0 looked like the obvious pick until it turned out to be API-only. Qwen-Image-2512 is the actual open model, and it runs.

Testing Qwen3.8-27B with DeepSeek Harness on a Radeon 780M. Large tasks stalled in reasoning; the goal tool let the agent split the work and continue.

Running MiniMax Music 3, an 8B/0.6B hierarchical autoregressive music model, on a Radeon 780M through ComfyUI. INT8 quantization to fit shared VRAM, hand-wiring the real node graph behind a collapsed UI subgraph, and a ROCm context-poisoning flake caught mid-generation.

Turning Craymel Ball's flat 2D sprites into usable 3D character models entirely on a Radeon 780M. TripoSR gets rejected, Hunyuan3D-2mini becomes the first usable tier, and multi-view input on Hunyuan3D-2mv fixes a measured depth-flattening problem that resolution and postprocessing couldn't touch.

Running LTX video generation on a Radeon 780M mini PC with no discrete GPU, through a VRAM ceiling, four LTX-2.5 loading bugs, a GGUF quant missing its architecture metadata, and a silently dropped boolean flag that turned coherent video into static noise.

A local codegen loop, escalating from qwen3:8b to qwen3.6:27b, reproduces its own past output, hits one real capability wall in a signed-zero bug, and ships eight tested functions into a working game.

AMD's own TheRock build now targets gfx1103 directly, so ComfyUI on a Radeon 780M doesn't need ZLUDA's DLL injection and security exclusions anymore, and it runs roughly twice as fast with the right attention setting.

ComfyUI wouldn't touch the GPU on this AMD chip at all. Getting real acceleration meant DirectML, then ZLUDA, and eight separate crashes before Stable Diffusion ran at real speed.

A small Python library that pulls point-in-time annual financials for Japanese stocks straight from EDINET's XBRL filings, so backtests stop seeing numbers investors couldn't have seen yet.

A multi-strategy screener for Japanese stocks that fetches into JSON and renders a static site, running end to end on a single 32-bit Raspberry Pi 2.

A quick look at how this site is built.

A fresh start for the blog.