We trained a character LoRA on Krea 2 and benched it against our production image stack. What ports, what breaks, the prompting rules that differ from Z-Image, and the LoRA lessons from 230 test renders.
We benched LTX-2.5 for two full rounds and got confident, wrong answers, because the official templates ship an LLM prompt rewriter that is on by default and buried inside a subgraph. Here is how to find it, how to actually turn it off, and what the model does once your prompts reach it intact.
The same edit three ways with a 4-step Lightning LoRA, Flux Kontext and full Qwen. What each trade-off actually costs.
One node takes three reference images. Person plus product, person plus scene, and the group shot - with measured timings.
One generation, seventeen outputs: the flat image, every object on transparency, and a free background plate.
We ran the same edit with and without a mask. The no-mask version won. Here’s when a mask helps and when it costs you.
Add canvas and let the model fill it. The node, the feathering, and the mistake that invents furniture.
The nine nodes behind every local instruction edit, what each does, and the one node that turns generating into editing.
Twelve generators, six prompts, one shot each, no cherry-picking, with the numbers nobody publishes: seconds per image, VRAM required, and the exact cloud price. Six of them run free on your own card.
Eros is a baked LTX-2.3 variant that generates synced audio and video from a prompt in one model. It’s wild, but it writes its own soundtrack, so it can’t dub your voice. Here’s the honest picture.