Fine-Tune by the Step, Diversify at Test Time: the Generative Stack Gets Cheaper to Steer

Training, priced by the step

fal put a meter on fine-tuning Wan 2.2. The text-to-video trainer costs $0.004 per step and the image-to-video trainer costs $0.005 per step, with a 100-step minimum — per-step billing rather than GPU-hours, which makes the cost of a training run knowable before you start. You upload a zip of ten or more images or clips, the trainer fine-tunes only the low-noise transformer of the Wan 2.2 A14B model, and out comes a .safetensors LoRA file you plug into the inference endpoint. At the stated rates, the cheapest possible training run works out to forty cents. For a creator, that changes the arithmetic: a custom style costs less than a coffee, and you can afford to re-run it every week. m:…5606

The inference side is metered too: the LoRA endpoint is billed at $0.10 per second. And the image-to-video trainer has an opinion about its inputs — it requires video training data and will reject image-only datasets. Cheap training, cheap inference, and a gatekeeper that forces you to feed it the right kind of data. Wan 2.2 generates 81 frames at 16 frames per second by default, and an auto-scale option rescales inputs to match that format. m:…5606

Or skip retraining entirely

In a preprint posted this week, the Mine, Sample, Render (MSR) framework proposes something cheaper still: do not retrain at all. MSR is a test-time approach that externalizes diversity into an explicit, corpus-derived sampling procedure, using the generative model purely as a conditional renderer of a sampled configuration. The authors argue that modern generative models struggle to sample diversely even when explicitly instructed to, and report that MSR improves both the diversity of generated outputs and their alignment with corpus distributions. They evaluated it across weather report generation, user simulation, webpage design, text-to-image generation, and agent-generated 3D assets. It is a preprint, so the results are the authors' own and should be read that way — but the direction is clear: steering is being pushed out of training and into the sampling step. m:…8826

At the heart of MSR is a mined schema: the framework extracts an interpretable categorical schema from a corpus of real artifacts, then samples from it before the model renders anything. The sampling logic is separated from the generative model entirely — the model's job is just to render the configuration it is handed. m:…8826

And a petaflop box to run it locally

The third piece landed in NVIDIA's newsroom: RTX Spark laptops are now on sale, offering up to a petaflop of local AI compute, and the company says it provides tested recipes and playbooks for running local AI on the machine. The implication for creators is a three-tier steering stack: rent per-step LoRA training when you need a custom style, use test-time sampling tricks when you do not want to train at all, and keep a local petaflop box for the iterations you do not want to meter. m:…9165

The local option is not one machine either: RTX Spark laptops are being built by ASUS, Dell, HP, Lenovo, Microsoft, and MSI. The creator stack now has an entry point at every level of spend, from a forty-cent training run to a $2,399 laptop. m:…9165