Unfortunately it can be a bit of a hit or miss if it actually gets the text correct, v2 is definitely better as was trained on more than double the images v1 was. I might look into trying to get it to be more consistent (it probably just needs even more training images)

Main tag: i-h43rt-5ubs

Helpful tags: {color} shirt

Example between the two versions

Prompt: masterpiece, amazing quality, best quality, ultra-detailed, 8K, ultra-detailed-eyes, newest, highres, absurdres, <lora:IHeartSubs:1> i-h43rt-5ubs, 1girl, solo, upper body, sitting, white shirt, SMALL breasts, colored hair, milf, white background, simple background

Negative prompt: realistic, young, makeup, tattoo, face markings, white eyes, white pupils, glowing eyes, eyeliner, bad quality, worst quality, worst detail, sketch, censored, artist name, signature, watermark, patreon username, patreon logo, window, low quality, 2koma, 4koma, signature, watermark, bad anatomy, jpeg artifacts, signature,

Steps: 15, Sampler: Euler a CFG++, Schedule type: Automatic, CFG scale: 1.5, Seed: 65849540634, Size: 940x1216, Model hash: 5d5352f5e0, Model: midnightIllustrious_mdntIllusBakedvaeV20, Clip skip: 2

How I use it:

Tool: Forge

Checkpoint: Midnight v20 baked vae

Steps: 20

CFG Scale: 1.3-2

Size: 940x1216 (or something similar)

I then img2img at 0.4 denoising strength, with the largest size to 2048. And then ADetailer the bejesus out of it (specifically with mask dilation around 40-60, inpainting mask blur around 30-50, and denoising around 0.3-0.6), the ass detailer is particularly useful for this one.