V1.2 changelog:
improved consistency
new dataset trained
V1.1 changelog:
train on 1.0 new dataset to integrate new linguistic depth and to add detail and precision
known issues: from version 1 onwards there are problems with hands and some anatomical parts. version 1.1 tries to fix some of these issues still present but less evident from v1
V1 changelog:
training with large dataset and prompt from llm, accepts a t5xxl style as well as the usual tags
best with : Deepshrink and Perturbed Attention Guidance
unfortunately to alter the clip and model to integrate some sort of t5xxl made it more unstable and needs more control over the generation so I recommend using comfyui and take example from the nodes from my images
V1 BETA2 changelog:
model completely redesigned from scratch
advanced training for fashion, photography and cinematic
merge DPO
beta features: lettering and magazine cover
trigger for magazine cover :
magazine cover
trigger for trained covers:
SURFACE
XIOX
DIOR
PAPER
ID
HERO
GQ
ADIDAS
VIPER
PORTER
HARRODS
LOVE WAN
V
INTERVIEW
THE FACE
TEENVOG
HARPER'S BAZAAR
ASOS
WONDERLAND
DAZED
MIXMAG
NUMERO
WOMAN
MOEVIR
ALLURE
WAD
ROSALIA
FASHION
COSMOPOLITAN
WALLPAPER
BILLBOARD
W
L'OFFICEL
GLAMOUR
VOGUE
FISHEYE
ELLE
TIME
VOUGE
PLAYBOY
P.S.: preferably use karras or AYS as scheduler
V0. 97b changelog:
Cinematic/sci-fi training 25 epochs 67000 steps
Fashion training 20 epochs 43000 steps
Detail + Unet training 10 epochs 70000 steps
+ DPO merge
CosXL versions :
Cosine-Continuous Stable Diffusion XL is a new stability.ai experimental SD model. The most notable feature of this schedule is its capacity to fix tonality issues with SD models and produce the full color range from 'pitch black' to 'pure white', performing much better than noise offset or LoRA solutions.
0.95c changelog:
more photographic, fix hands, fix eyes
close-up photography
better skin details
better raytracing
finetuned on new dataset
0.9 changelog:
more graphic (added 4 trained loRAs)
finetuned on best results of 0.75
0.75 changelog:
more sci-fi
more cyberpunk
more cinematic
Introducing SDXL HK 0.9, the versatile amalgamation of cutting-edge advancements! This iteration pushes the boundaries of realism and sci-fi, evolving into a powerhouse of innovation. What's new in this version?
Enhanced Graphics: We've amplified the visual experience by incorporating seven meticulously trained LoRAs, enriching the model's graphic capabilities to create more immersive content.
Fine-tuning: Building on the success of the previous version (0.75)
What set apart SDXL HK 0.75?
Sci-Fi Infusion: Infusing more elements of science fiction, the model transcends traditional boundaries, weaving intricate and compelling narratives into its creations.
Cyberpunk Vibe: Embracing the essence of cyberpunk, this version delivers a futuristic, edgy feel, opening doors to a plethora of creative possibilities.
Cinematic: Elevating the cinematic experience, the model crafts visuals that resonate with cinematic grandeur, bringing stories to life in an immersive manner.
Merged Model: SDXL HK 0.75 was built on a model by Afroman4peace. Merges base, refiner, some LorRAs, and the model by Afroman4peace. Additionally, it underwent a second merge with two finely-tuned LoRA models specialized in photographic content.
Maintaining Consistency: The primary goal behind this model's creation is to uphold a consistent quality standard alongside my LoRA models, ensuring coherence and reliability across various creative endeavors.
Description
FAQ
Comments (16)
I really like your models, probably one of the best for creativity and non-conventional art!
I know sd3.5 medium is underwhelming for realism but it is very creative for painting and art, and I wonder what would sd 3.5 medium HK would look like....
🙌🔥🙏Thank you! There's a lot of work to do on the medium, but I might give it a try
Yes the model seems to be a messy hell, but someone with aesthetic knowlege may guide it in the right direction, at least for experimental purposes.
@Primaveri you're right, I'll do some experiments
Nice update! Do you consider making a CosXL version out of this?
Thank you! right, I could also upgrade to the cosxl version
@Ciro_Negrogni That would be great! Love your 0.97 CosXL model!
I like and don't like the results. Its really trained, so that's good, the results I get with the same prompts are different from other models. But everything is on the dark side and it repeats. People are cut in half or I get triplets when I just wanted one thing.
i get a lot of AI look, where bottles are squished or lines are soft, like using steampunk plumbing parts. I get an interesting creation, but up close its a bit of a mish mash. I find the details in many of these are small, like If I wanted a clear tank, it will make a bunch of smaller ones. I'll try this in different modes to see what might work best.
this version of xl is very special i recommend using deepshrink and perturbed attention guidance
@Ciro_Negrogni i don't know what those things are. I have a very simple AI thing, they never updated it but I liked the interface, easy diffusion.
@frankmike for easydiffusion I suggest you try version 0.97b of the model. unfortunately to alter the clip and model to integrate some sort of t5xxl made it more unstable and needs more control over the generation. for version 1 i recommend using comfyui and take example from the nodes easily copied from my images
It would be so nice if you share a comfyui workflow with a minimal set of components and comments to run the model
good idea, I'll create it as soon as I can
@Ciro_Negrogni any news?
@Olbanets https://civitai.com/articles/9218
cant download via api in google collab already put the token , anyone can help