———— ▬▬▬▬ ———— ▬▬▬▬ ———— ▬▬▬▬ ————

❗ March 26th - NOISE OFFSET v2 UPDATE - Darker & Better quality

———— ▬▬▬▬ ———— ▬▬▬▬ ———— ▬▬▬▬ ————

DIABOLIQUE MERGES - Version Preview

❗ ORIGINAL - First /original merge. Semi-realistic to realistic. All the other versions below are its babies. This will stay as it is for reference.

❗ NOISE OFFSET - Merge with noise offset added to it. Recipe is photoreal + fantasy.

❗ PHOTOREAL - Further merged for photo realism.

❗ FANTASY - Strong fantasy prompting capabilities. Can do semi-realistic /digital /anime. Offspring of original merge above.

❗ ANIME - Anime merge is an offspring of the fantasy merge. Aimed towards anime, does digital arts /fantasy as well, decent at it, with right prompts.

❗ MANGA - Manga merge is an offspring of anime merge. Primarily aimed for manga /illustrations, anime and such.

———— ▬▬▬▬ ———— ▬▬▬▬ ———— ▬▬▬▬ ————

ℹ️ PROMPT BOOK v2 (Download the PDF) or CHECK ONLINE (Excluded NSFW from this, separate book with NSFW coming later)

ℹ️ IMGUR ALBUM - Fun experiment for everyone to try. Prompting w lyrics! Feed SD just with lyrics. Here is an example album - images generated with few lines from Iced Earth, metal band.

———— ▬▬▬▬ ———— ▬▬▬▬ ———— ▬▬▬▬ ————

ℹ️ What went into making of all the models

QGO-10b /F222 /BasilticAbyssDream /Wonder /Refined /Clarity /HASDX /RPG /HARDBlend /Hunter 69 /Dreamshaper /Cyber Realistic /Deliberate /Elegance /Elysium /Galaxy 3 /Generic Female /Project AIO /LOFI /Realism Engine /rMad /Woop Photo /A to Zovya /WLOP /NED /Perfect Sketchbook /Dark Sushi /Epic Illustration /iCoMix /Comi Noir /Counterfeit /Meina /ReV Animated /Tri Phaze /EZ (Seems to be gone)

There were a lot of merges in the process and its kind of hard to tell the exact recipe at this point. Big thanks to all the authors of the merges above for their amazing work.

———— ▬▬▬▬ ———— ▬▬▬▬ ———— ▬▬▬▬ ————

ℹ️ Which model should you chose? Quick read on these

Pruned models & their differences

Fp32 /Fp16 & their differences

ema-only weights use less VRAM - suitable for inference

ema /no-ema weights use more VRAM - suitable for fine-tuning

———— ▬▬▬▬ ———— ▬▬▬▬ ———— ▬▬▬▬ ————

My workflow

I run ✔️ hires fix with R-ESRGAN 4+ for photoreal /realistic outputs.

I run ✔️ hires fix withR-ESRGAN 4+ Anime6B for anime /manga outputs.

I run ✔️ hires fix with 4x-UltraSharp at times for various tasks.

CFG scale /steps vary per model /desired output, so does the sampler. Experiment :)