Custom nodes (works exclusively with this in comfy): https://civarchive.com/models/2087992?modelVersionId=2362403
ONLINE TRY: https://www.seaart.ai/workFlowAppDetail/d4jcrmte878c73dpsqr0
ANIMAX v3: A System, Not Just a Model
Welcome to ANIMAX v3! This is not just another model release; it's a complete, precisely-tuned creative system, born to achieve the highest quality and establish artistic consistency.
The Core of the System: The Container Model
The heart of the system is a unique .safetensors container model. This file is a fusion of a carefully selected base model and two complementary, specially-trained, large-scale LoRAs. The result is an incredibly powerful and coherent foundation that produces stunning images on its own.
The Unique Advantage: The "LoRA Enhancer" Capability
What makes the ANIMAX system truly special is how it interacts with other LoRAs. Because our two main LoRAs are already an integral part of the model, they create a highly stable and receptive base. As a result:
When you load an additional, standard LoRA into the workflow, the ANIMAX model enhances its stylistic traits, leading to a final image that is sharper and more integrated than if you were using a traditional model.
Technical Specs and Requirements
Native Resolution: 1536x1536 pixels.
Required Tools: To use the system, it is essential to install our custom ANIMAX node pack. This pack includes the Artisan Engine, which is the "key" to unlocking the power within the container model.
Recommended Usage: For the best results, load the provided workflow (.json) file.
A Note on Training
Many may ask how the model was made. The training process has extremely high VRAM requirements (48GB+) and uses specialized techniques, so the exact methodology is not public at this time.
Join the Development!
This system is built on community feedback. Please share your experiences, your tests, and your generated images with us in the comment section. Every piece of feedback helps make ANIMAX even better.
Description
FAQ
Comments (11)
lie! it can't do full hd, i set it to 1080x1920 and my computer freezes completely, or gpu out of memory error, i have gtx 1070ti 1336x1640 is the max it can do and any sdxl model can do it
8gb vram is not enough for full hd resolution, you need at least 14gb
BUY Better GPU and try again
It works for me, though you might want to use this with refiner when going 4K/8K
Skill issue. I can do 1920x1080 on my 4GB GPU. ComfyUI have VAE Decode (Tiled) - it's slow, but it actually can process big latent space on a potato setup.
good model. can you tell please best sampler and scheduler and cfg rate for best coherence\aestethics level?
Of course , my Sampler: DPM++ 2M SDE, Schedule type: SGM Uniform, CFG scale: 5, Steps: 70-150, and the resolution varies but my favorite which is fast and sharp is Size: 1288x1624, I like this one the best but obviously it's up to the person
but if you have the patience I recommend the 1440x1920 resolution but with less steps like 40/50
Another large fine-tune based on SDXL. I can highlight excellent anatomy. Eyes and fingers always come out great at a resolution of 1568x1568. The model demonstrates quite solid knowledge overall. I tested several NSFW concepts, and the results were impressive. I don't see the point in thoroughly testing SFW content since that usually never presents any problems.
The model has a very strong default style, which is hard to change significantly using artist-based prompts. As mentioned by the author, additional unwanted characters often appear in the background,and sometimes not even in the background. This issue happens quite frequently, especially with less popular concepts.
The SCORE system is an interesting concept. It's a bit unusual to see many weights after tags, but one can get used to it. Some people are probably already familiar with it. In general, this is quite good for a first version.
I’d also like to ask. Why did you choose the SDXL architecture specifically? Given the computing resources you likely have available, it might have been possible to try out a newer, more powerful architecture. I'm absolutely not criticizing your choice, just genuinely curious.
First of all, thank you for testing the model in such detail: )
as you mentioned it wasn't a technical obstacle to training but a lack of budget, unfortunately renting such a card is already horibile per hour let alone 6pcs, it's no secret that training the model only costs us $1075, so thinking about a more modern technology training a basic model would have cost 3-4 times more
although we were teasing with the flux model, we did a lite version but it's so hard to get around the NSFW filters in the text encoder that we were stuck there, we tried many mods even the full retraining but the result was horrendous so we gave up, then the team split up and I was left alone and my personal experience is with SDXL so it was understandable that the work I started was moved there,
Amazing model! Thanx for share!!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.