Myths and Facts

Myths and facts

The usual advice on brand LoRAs, checked against papers, model cards and trainer source on 2026-10-15. The sources are in the repository's research note.

Advice	What holds	What brandwright does
Train two LoRAs with distinct trigger words, and put both words in the prompt to switch them on	A LoRA acts whenever it is loaded, and its weight is the switch. A trigger only conditions it.	It keeps a trigger per LoRA, but trains the site LoRA on top of the brand LoRA, because sub-site screenshots are brand screenshots too.
1-2 LoRAs is the sweet spot; 5 or more turns text into hieroglyphics	The direction is right: merging more LoRAs loses detail. The thresholds are folklore.	Two LoRAs are enough when the second is residual.
Keep the weights summing to about 1.0	It is a heuristic for LoRAs that overlap. diffusers' own example uses 0.7 + 0.8.	Brand 1.0 + residual site 1.0, then a sweep.
Use a UI-specific base model	None is in mainstream use, and the academic ones are old and non-commercial.	Z-Image, Qwen-Image-2512 or FLUX.2 klein 4B, chosen for text and licence.
Image models cannot do small text, so a designer must rebuild everything in Figma	Small text and icons are still unreliable. Qwen-Image-2512 gets one to three large words right about 70% of the time, and sentences under 40%.	The code route builds the final mock-up as real HTML.
Train in an enterprise cloud, because you cannot train on the payment site	Training never happens on a site, and a 16 GB card trains these models locally.	Local by default. The real limits are personal data and company policy.