Welcome to AIcAI - AnimeHaven✨
I'm back at it again !
These weeks were rough... I lost my workflow (it got corrupted by a severe PC crash while generating an image) and all my custom nodes. I wasn't ready to make a new one, thus the delay for the new model. Sorry for that ! But fear not, as I'm about to unveil my greatest creation as of today !
Oh and by the way, I have a suprise for you all:
I HAVE PUBLISHED MY CUSTOM WORKFLOWS AND NODES TO GITHUB !!!
Custom nodes: Alcai-nodes
Image Generation Workflow: workflow T2I
Upscaling Workflow: HiresFix
(Click on the download button for the workflows and add them to the user\default\workflows directory inside your ComfyUI folder)
Thanks for reading and enjoy this new model !
(Real AlcAI is out, now time for ALcAI BOT to enter the stage.)
The journey continues. After pushing the Awakening series to its limits, I felt it was time for a new direction, a new philosophy. Today, that new chapter begins.
A new series is born: Elysium, a model crafted for the ultimate in artistic polish and vibrant expression.
!!! CRITICAL - CUSTOM VAE REQUIRED !!! This model was trained with a custom VAE for the best results. Please download and use it: Anzhc's SDXL Anime VAE (B3)
AIcAI - AnimeHaven - Elysium V10!
Elysium isn't just an upgrade; it's a complete evolution. Built on all the experience from the Genesis and Awakening series, this model is fine-tuned to deliver images that feel like they're straight out of a professional artbook or a top-tier gacha game. It's all about clean, vibrant, and perfectly composed illustrations.
So, what makes Elysium the new benchmark?
🎨 Art Direction This is the core philosophy of Elysium. It has a masterful command of a clean, high-impact commercial art style. Generations feel incredibly polished, with perfect lighting, vibrant colors.
🌟 Vibrant, High-Energy Colors & Contrast: Elysium takes the color science of Awakening and cranks it to eleven. Colors are exceptionally rich and saturated, and the model excels at high-contrast compositions, including stunning selective-color effects on monochrome bases. Your images will absolutely pop.
🎬 Creative & Confident Compositions: This model thinks like an artist. It has an even better grasp of dynamic framing, creative layouts, and telling a story through a single image. It handles elements like ornate borders and minimalist backgrounds with equal confidence.
✨ The Apex of Detail & Clarity: Everything you loved about the previous models is here, but sharper. The linework is flawless, the rendering of complex outfits is impeccable, and the multi-character capabilities are as strong as ever. It's the pinnacle of clean execution.
Good to Know (The "Catch"): Elysium is a model with a strong, opinionated art style. It is designed to produce bold, clean, and vibrant illustrations. If you're aiming for a softer, more painterly, or gritty look, it might require more specific prompting to pull it away from its core strengths. But for that high-energy, professional anime aesthetic, Elysium is in a league of its own.
Recommended Settings (Generally consistent across all XL models):
Dimensions: Start around 1024x1024. Handles rectangular aspect ratios well. Can push to 1536x1536 and even higher with adequate VRAM.
Steps: 25 to 50
CFG Scale: 4.0 to 7.0
Sampler:
Euler_a,DPMPP_2M,DPMPP_2M_SDE,DeisScheduler:
Karras,Sgm_uniform,betaHires. Fix: Still recommended for pushing detail or resolution further.
Hires steps: 15 - 30
Denoising strength: 0.3 - 0.7
Notes:
Excellent for recreating specific character styles or artistic vibes. Use artist names or series titles in prompts for style emulation.
Releases notes from previous versions:
Elysium Series (IllustriousXL 2.0 Base - FP16):
Elysium V10: The debut of a new series focused on the ultimate artistic polish. Built for 🎨 Professional Art Direction, 🌟 Vibrant Colors, 🎬 Confident Compositions, and ✨ The Apex of Detail.
Awakening Series (IllustriousXL 2.0 Base - FP16):
Awakening V30: An evolution that enhanced 🧑🤝🧑 Multi-Character Management, 🌧️ Atmospheric Depth, and 🎨 Stylistic Fidelity, bringing it closer to the Genesis series' moodiness while retaining XL clarity.
Awakening V20: A masterclass in refinement. Introduced the game-changing ability to create character sheets with unprecedented consistency, perfected intricate detail rendering, and fine-tuned lighting dynamics for flawless execution.
Awakening V10: Marked the beginning of the "Awakening" era. A quantum leap in quality with: 🎨 Elevated Artistic Direction & Lighting, 😊 Enhanced Character Nuance, 🖼️ Masterful Composition, and 🔞 Exceptional NSFW Coherence.
Ascension Series (IllustriousXL 2.0 Base - FP32 CLIP/VAE):
Ascension V30: Took a major step forward by significantly enhancing multi-character interactions, increasing atmospheric depth, and boasting superior compositional strength.
Ascension V20: A refinement step that aimed for enhanced vibrancy, consistency, and character appeal, while adding robust NSFW capabilities.
Ascension V10: Marked the first chapter on the IllustriousXL 2.0 base. Focused on Stylistic Mastery & Clarity, with unparalleled style versatility and defined linework.
Description
This model includes the following:
Better textures
Better lighting
Better stylization
FAQ
Comments (10)
Why its over 8gb? are you using fp16 unet+ fp32 clip?
Yes but not sure that it is the reason why its over 8Gb
EDIT: It is the reason why its over 8Gb
Fot those wondering how I can possibly that fast with my merge, my workflow has been optimized a lot since I've begun using comfyUI (back in 2023) but I didn't have a powerful enough GPU at that time (A 3060 laptop with NO VRAM 6GB of VRAM).
Now Everything have changed: I have a 4080 SUPER 16GB VRAM (on my way to get a 5090). Also, I've made custom nodes to help me produce these models.
I might be envisaging creating a github repo and let the community do its way with what solutions I've come with, but first I might need to refactor the code a bit (you know, AI and all ...)
Hope you'll stay all with me with each release :)
Hi, could you elaborate on >> THE CLIP AND VAE ARE USING FP32 ! If you don't use that, result might be vastly different !
Is there anything special that has to be done for local installs (I'm using reforge)??
Add these Arguments: "--vae-in-fp32 --clip-in-fp32"
A quick tour on the github repo for reForge got me the answer (btw I don't use anything else other than ComfyUI so I can't tell you where to put the args, sorry)
@alcatraz974 I know where they will go. As an aside would that be why the intial images made seemed "zoomed in" or is it because IL 1.1 was trained on larger images?
@AI_Yaya_2024 I don't know. There's a lot of random things that can be the cause of your "zoomed in" feel. It can be the prompt, it can be the merge... Can't really tell you what to do about that. And generally speaking, anime images are, a lot of time, close-ups.
@alcatraz974 it was testing a prompt (no change) from an earlier job with another model. I changed the "canvas size" form 832x1216 to 1024x1536. I guess that's the caveat with IL 1.1 being trained on larger images.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.










