Things I build while running models on my own hardware, cleaned up enough for other people to use.
Every section stands on its own, with its own installer and its own docs. There is no shared setup to get through first.
Video with sound, generated at home. MiniMax H3 writes the picture and the synchronized audio in one pass, so voice, effects and music come out of the same generation as the frames. Open weights.
Thirteen ComfyUI workflows in two sets. One set runs on any single 24 GB card. The other puts every card in a multi-GPU box to work and hits 3.12x on four cards, taking a 15-second shot from 19 minutes of sampling down to 6:41.
Ships with the custom nodes, the patch that makes the turbo LoRA actually load, and a three-command installer.
cd minimax-h3 && ./install/setup.sh checkImage generation, editing and transparent PNGs on one GPU. The baseline is an RTX 3090 with 24 GB VRAM, using official INT8 ConvRot weights.
Five native ComfyUI workflows, an isolated installer, SHA-256 model verification, and a script that submits a generation and checks the resulting PNG. Includes an English guide, a Russian guide, and an installation brief for an agent.
For smaller GPUs, use the low-VRAM guide: GGUF Q4_K_M, CPU offloading, 7 workflows, and measured memory requirements.
cd qwen-image-2.1
python3 install/setup.py checkAgent skills that stand on their own.
hand-drawn-canvas-animation makes short films with authored poses and expressive strokes in JavaScript and Canvas 2D. It covers pencil, ink, risograph, screen print and drawings interacting with real photos, plus sand animation and projected paper. The 60-second phoenix film combines the materials through one story, including a moving storm and wings that unfold from a book. The skill includes editable examples, original sound, an MP4 renderer and browser regression checks. No Blender or video-generation model is required.
cp -r skills/hand-drawn-canvas-animation ~/.agents/skills/More lands here as I clean it up.
4x RTX 3090, 96 GB VRAM total, PCIe 3.0 x16, no NVLink, 320 W per card. Ubuntu.
Numbers in these READMEs come from real runs on that machine, not from datasheets. Your absolute times will differ, the ratios usually hold.
There is one thing worth checking on your own box before anything else. Cards ship at whatever power limit the vendor set, and it is often well under the default. Mine were at 220 W against a 350 W default, and raising them to 320 took an identical run from 220 seconds down to 170.
nvidia-smi --query-gpu=name,power.limit,power.default_limit --format=csvI post the benchmarks behind all of this. Local inference on 4x RTX 3090, open-weights models, argv and raw logs attached.
MIT. Take it, change it, ship it.