V1.5 just released. The writing is more consistent although not quite at the level of FLUX.1, which still leads.

With TrueToLifeSDXL-XXX-Lightning, I'm trying to develop a model that's both anatomically accurate (no more bad hands) and can add selective text to an image without it looking like an alien wrote it. While it can be used for a general purpose model, I think its strength lies in the way it mimics real-world analog film photo's. The current version seems to be working ok at the moment but I feel there could be some improvements in getting longer more coherent text to render.

With the introduction of FLUX.1 to the scene, more effort is going to be spent there. V1.4 is a more artistic, general model, without the text generation capabilities of earlier versions. FLUX.1 has made further experimentation with text pretty much unnecessary. V1.5 is more realistic again, with better handling of anatomy. Next challenge is to fine-tune FLUX Schnell.


General generation parameters for A1111:

DPM++ SDE, SGM Uniform, 5-7 steps, CFG 1-2, 0.3-0.45 denoising.
Euler A, SGM Uniform, 8 steps, CFG 1-2, 0.3-0.7 denoising.
Ultrasharp and Foolhardy_Remacri upscalers work best, but YMMV.

It works best at 512x768, 2x upscale, 2:3 or 3:2 ratios, but other sizes work as well, key is to test and see what works for you.

Generic prompt format:

<style> <subject> <location> <text (one or two words work best)> <smaller less important detail>

Example:
<70's style photograph> of a <40yo middle-aged graying brown haired woman standing> in an <abandoned nightclub>, there's a <green neon sign saying "OPEN":1.3 on the wall above her head>, <she looks tired and worn out>.

A negative prompt is usually not needed, but if you find things showing up that you don't want, usually a single word or two is enough to suppress unwanted items.