.jpeg)
--This notes article is a continuation of my first ILXL Explorations Article, I'll be noting here the current state of evolution the this amazing model branch--
--if I glossed over anything or you have more info to add tell me in the comments and I'll add it--
the year 2025 have been quite the amazing year for visual AI with a lot of open source heavy hitter model entering the scene, it its safe to say that SDXL (🗓Jul 26, 2023) is still going going strong for yet another year all of that though the amazing model branch Illustrious-XL v0.1 (🗓Sep 25, 2024) pushing that old model to preform some great feats ..
I mean lets be real here .. what's last time you had worry about body horror or mutated hand ? or iterating over a prompt hundreds of times to get what you want ? even better it has tons of styles out of the box to mix and mash from.. that's by itself is so damn impressive in my book!
Checkpoints Family Tree:
--we came a long way since the original ILXL with each sub-branch/finetune/Merge bring something to the table, so here's what I tested so far--
Base SDXL (🗓Jul 2023)
Kohaku XL Beta 5 (🗓Nov 2024)
Illustrious-XL v0.1 (🗓Sep 2024)
(📚Danbooru Dataset 2023)
(📌brings a better prompt control then PonyXLV6)
(📌brings a wide range of per-learned Artists Styles and Characters)
(📌easier to train on then PonyXLV6 and Base SDXL with far better success rate)
.Illustrious-XL SmoothFT (🗓Sep 2024)
(💚my go-to model for LoRA training)
(📈less overall artifacts)
(📈a slight color improvement over the base)
.NTR MIX (🗓Nov 2024) (🍹Merge)
(📈Default Style)
(📈Vivid colors)
(📈Handle reflections and glowing well)
(📉Less accurate artists styles)
(📉Less LoRA friendly)
.NoobAI-XL V-PRED 1.0 (🗓Dec 2024)
(📚Danbooru Dataset 2024)
(📚E621 Dataset 2024)
(📈wider color range especially darker and brighter ones)
(📈great artists styles accuracy)
(📈wider tags knowledge)
(📉prone to anatomy errors but less then the base model)
(📉less deterministic CFG)
.Plant Milk - Hemp II (🗓Apr 2025) (🍹Merge)
(📈Richer Details)
(📈Better Perspective Gens)
(📈Default Style)
(📉Less accurate artists styles)
.WAI-illustrious v14 (🗓May 2025) 🏆
(💚 the swiss army knife of Illust XL !)
(📚Danbooru Dataset mid 2024 ??.. i think)
(📈non-intrusive default style)
(📈great anatomy accuracy)
(📈great artists styles accuracy and quality)
(📈better prompt control)
(📈extremely LoRA friendly)(📈works well with most known samplers)
(📉struggles a bit with brightness control, specially dark gens)
.RouWei 0.8 V-PRED (🗓Jun 2025)
(📚Danbooru Dataset Apr 2025)
(📈more artists styles )
(📈great color range)
(📉needs BREAK to better use artists styles)
(📉can get prompt heavy sometimes to get what you want)
(📉embedded styles sometimes bleeds into characters or concepts )
.
Illustrious-XL 1.0 (🗓Mar 2025)
(📚Danbooru Dataset Apr 2025)
(📌brings a higher native resolution of 1536×1536)
(📌somewhat compatible with LoRAs trained on v0.1)
(📉while better then base v0.1 its really underperforming compared to fintunes from 2024)
(📉trained LoRAs on this model had more issues compared to old one, the unusual new native resolution might be what's causing that)
(🎭sparked a lot of drama following its announcement leading to future Illustrious-XL models releases like ILXL V2 and especially ILXL V3)
Gen Tests:
the following image generation test will mostly use the same settings across different checkpoints to illustrate differences/improvements
V-pred models will use EulerA, 25 steps, 4 CFG
EPS models will use EulerA, 25 steps, 6 CFG
Tests:
[Base 0.1, SmoothFT, NTR4, NAI1.0, WAI14, PlantMilk-Hemp2, RouWei 0.8]
[Base 0.1, SmoothFT, NTR4, NAI1.0, WAI14, PlantMilk-Hemp2, RouWei 0.8]
[Base 0.1, SmoothFT, NTR4, NAI1.0, WAI14, PlantMilk-Hemp2, RouWei 0.8]
Shaded Style - Artist Sakimichan
[Base 0.1, SmoothFT, NTR4, NAI1.0, WAI14, PlantMilk-Hemp2, RouWei 0.8]