About Us – The LoRA Test Kitchen
I am Mara Vance. I live in Richmond, Virginia, with my husband Daniel and our seven-year-old daughter, Iris. By day I work remotely as a product-design consultant, the kind of work that forces you to watch how people actually use a system rather than how the slide deck claims they will. By night the garage becomes a compact AI lab: a workstation, a camera corner, storage shelves, and a whiteboard covered in experiment notes. That garage is where The LoRA Test Kitchen started.
I did not arrive here through a formal machine-learning degree. I arrived by treating generative models the same way I treat software prototypes: define the target, change one variable, record the result, and repeat the test. After years of documenting user behavior and workflow friction, the habit transferred cleanly. A pretty generation is interesting. A workflow that still works on Monday morning is useful.
The Garage Lab and the Paper Notebook
The lab is not glamorous. It is a converted corner of the garage with enough power, enough cooling, and enough shelf space to keep datasets and failed runs from disappearing into the void. I keep a handwritten “failed experiments” notebook and refuse to delete a bad result before writing down what caused it. The notebook is older than most of the models I test. It sits next to the keyboard the same way a lab journal sits next to a bench.
Daniel is a civil engineer. He understands documentation the way I understand iteration. Iris is seven and already knows that “Mom is running a test” means the garage door stays closed and the whiteboard stays covered until the numbers are written down. The family rhythm is ordinary: school drop-off, remote client calls, evening baking, occasional film photography on the weekends, and the quiet satisfaction of repairing a vintage radio that still works after sixty years. Those slower hobbies shape how I evaluate AI. A camera that only produces one lucky frame is not reliable. A radio that drifts after an hour is not finished. The same standard applies to a LoRA that looks perfect in training and collapses the moment the scene changes.
Learning What Actually Moves the Output
My early LoRA attempts followed the usual pattern. I trained something that looked excellent on the validation set, celebrated for an evening, then watched it fail in every new context. The failure was instructive. Dataset size, caption quality, weight strength, and the difference between “looks good once” and “holds up across ten seeds” became concrete instead of abstract. I started measuring more carefully: model and LoRA versions, exact prompts, dataset details, training and inference settings, hardware, runtime, cost, failure rate, and limitations. Those records are the backbone of every major post on this site.
I focus on MiniMax H3 Turbo because it is the model I actually run and re-test. I do not write about features I have not used. When a workflow claims to be production-ready after one successful generation, I treat that claim as a hypothesis to disprove. The useful question is rarely “Can it look good?” The useful question is “What breaks when I change only one variable, and is the remaining process still worth the cost and the friction?”
Why the Test Kitchen Exists
I write for AI creators, indie developers, technical artists, and small creative teams who need practical results rather than marketing claims. The site sits between glossy demo reels and dense research papers. Every major experiment records the settings so another person can reproduce the run or deliberately break it. Successful results and failed results both belong here. A pretty demo is not a repeatable result.
The five sections—The H3 Bench, LoRA Workshop, Workflow Kitchen, Failure Files, and Field Notes—mirror the way the work actually happens in the garage. Benchmark the model. Build and diagnose the LoRA. Assemble a full recipe. Document the failures. Step back and notice what the process is teaching. The signature rule is simple: run the test, save the settings, show the failure. If a workflow cannot survive that sequence, it does not belong in the permanent shelf of recipes I keep.
I am not selling effortless automation or guaranteed commercial outcomes. I am offering the same discipline I use on client workflows and on the whiteboard in the garage: change one thing, measure what happened, write it down, and decide whether the result is still worth keeping on Monday.
— Mara Vance
The LoRA Test Kitchen
Richmond, Virginia