<h3 id="upcoming-changes-(truetolifesdxl-xxx-lightning)-387xvxrij">Upcoming changes (<a target="_blank" rel="ugc" href="https://civarchive.com/models/568507/truetolifesdxl-xxx-lightning">TrueToLifeSDXL-XXX-Lightning</a>)</h3>I've been working on the text understanding and decided to try a new approach to model merges and fine tuning for understanding and anatomical accuracy (almost, but not quite there on the hands...). So far testing has been rather positive and while the photography side seems sorted, I'm trying to generalise more. New model has just been published. Prompt: 35yo woman wearing sheer white nighty, auburn hair, green eyes, freckles, standing in the bedroom, (holding a (((handwritten))) sign)( saying "Really NOT SD3":3) <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/113b028e-3384-4229-84ac-240cb917df04/width=525/113b028e-3384-4229-84ac-240cb917df04.jpeg" /> Changes (v.3.5): Refining the model to accept shorter natural language prompts, but still accept some negative prompts to steer it more in the direction you want it to go. Also continuously refining the image clarity and coherent text output.A: Better text. B: Better hands. C: Merge of A &amp; B, WIP.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c20bacec-98a3-44ac-b5b8-715785f188c5/width=525/c20bacec-98a3-44ac-b5b8-715785f188c5.jpeg" />v2.0-Final - v3.0 changes: Just uploaded a new checkpoint with a few more tweaks to get better details. I still get good results with the following settings: DPM++ SDE Uniform, CFG 1, 5 steps, Hires sampler 4x_Ultrasharp/4x_Foolhardy_Remacri at 0.3-0.41 denoising, 5 steps OR DPM++ SDE SGM Uniform, CFG 1, 6-7 steps, Hires sampler 4x_Ultrasharp/4x_Foolhardy_Remacri at 0.3-0.41 denoising, 6-7 steps As usual, prompt examples in the image, shows adherence to the prompt. Short descriptive positive prompts with 4-5 key elements work best, with 1-2 negative prompts to get rid of unwanted elements. v1.0 Re-upload of Rev.3 of the model, as clip merge was corrupted, resulting in issues on ComfyUI. This is a merge of <a target="_blank" rel="ugc" href="https://civarchive.com/models/383364?modelVersionId=453382">CreaPrompt_Lightning 1.3</a>, and <a target="_blank" rel="ugc" href="https://civarchive.com/models/43977/leosams-helloworld-xl">LEOSAM's HelloWorld XL.</a> Two of the nicest models to work with, imho. Additional LoRA's have been merged to make it more compliant with text and generate hands slightly better. It does sometimes generate NSFW images, so add nsfw to the negative prompt to control it. Test images show the prompt following and writing adherence. No negative prompts have been used to show the baseline capabilities. PS. To get better spelling, try enclosing the words in brackets and weighting the phrase to try force it to co-operate. eg. (woman holding a sign saying "Test Me":1.3)

TrueToLifeSDXL-Lightning

Continued refinement. This should work without a VAE (the VAE is baked in), but if you have any issues, use the sdxl fp16 fixed one, that works on Comfy and A1111.

Files with SHA256: b72f338842483c14dc3c60bc56457f117f8ebf56b188aab7f6941aa2398e9b02