Fiz-Rot - LyCORIS LoCon - NoobAI-XL Epsilon Pred 1.0 - InProgress版本2.4 - 000029 (ID: 1072812)

This is an in-progress training as I am learning how the Illustrious (Or NoobAI, I honestly don't know anymore) family of models work, feedback is appreciated

LoCon trained on Fiz-Rot's artstyle for NoobAI EP 1.0 Version: https://civitai.com/models/833294?modelVersionId=1022833

I'm thinking about retraining for the VPred version eventually and it will likely be an update to this page

Current known issues with the LoRA:

• Hairclips are generally unusable in most capacities, if they show up randomly (usually on black hair or swept bangs) put "cross" and/or "hairclip" in negatives

• Pupils are rarely present, eyes will tend to be textured with implied pupils to unicolor solid with a light glow

• Nose tends to be dot nose when I would prefer the button/snub nose

Issues that (possibly) stem from the base itself:

• Both inconsistent and volatile, it's highly recommended to mix this lora with inbuilt artist tags to get some consistency

• (As per above) The style without artists ranges from 2.5d to 2d, the use of "realistic" and/or mid-weighted "3d" can lean more in to the 2.5d aspects of his style

• If you want to lean more into the traditional aspects, the use of "cross hatching, traditional media" emphasizes these parts

Thus far my most successful permanent positive/negative combo with this model is:

masterpiece, best quality, good quality  

lowres, worst quality, low quality, oldest, jpeg artifacts, blurred, artist name, signature

However there are likely better combinations and more experimentation is needed

General Workflow for Sample Images (on ComfyUI):

• Sampler from images is done with Restarts with segments of [3,2,0.06,0.30], which are values from the paper on Restarts

  • Note: Using restarts tends to smooth out the image more but makes it more coherent, your mileage may vary and it's up to you if you want to use them

• No special scheduler is used

• Conditioning is split (A1111 BREAK) into 3 parts, generally BG and camera/Subject Face and attributes/Subject body and clothing

• Initial resolution is selected from the SDXL whitepaper appendix I: https://arxiv.org/pdf/2307.01952

• Image is upscaled by 1.5x using Lanczos resampling

• Second sample pass is the same, but at around 20 steps at .4 denoise

• A facedetailer is used after the upscale

Recommendations:

• Strength can be any value at or below 1

• Since it's trained on the NoobAI EP1 version, it's expected to only work on that model or merges of that model. This has thus far only been tested on NoobAI. It's unlikely to work on other Illustrious finetunes.

• You are highly encouraged to use artist tags alongside the LoRA, while it is usable without them the model really shines when artist tags are used.

描述:

Initial decent version

Non-permanent, highly likely to be archived in the future

训练词语: