NEW NODE PACK FOR COMFY:

https://civarchive.com/articles/24121/z-fuse-comfyui-node

This new node library is something I made to make better use of merging/using loras for Z-Image. It allows users adjust lora strengths by block layers while showing how much influence loras have for each block. No more messy mangled gens from using a chain of loras. Instead of adding strength it remakes the layers and adjusts accordingly without over saturating layers allowing you to have more control over how loras behave in a stack.

DE-TURBO:

A de-turbo model has been released making it easier to train new Loras. So hopefully my next release will be the final version. Either that or I will merge my loras into the de-turbo and finetune the model and release that, i'll probably do both.

Activation Tags: (Additional Trigger Words Down Below)

uwu_clarity <-- (Included an activator to help steer the LoRa, although it's not exclusively needed it may help.)

pornographic <-- (For the gooners and coomers. You may need additional LoRa pairing for specific positions.)

v4.0:

I'll be honest, training this model has been ROUGH. I will probably pause here at v4 and wait for the full model to release. Unless I find a new training method. While I can generally get a style or composition I can't for the life of me make spicy content work reliably. SO this v4 will again be a lightweight small Lora.

v3.0:

This is a LARGE retrain + a merge with the other Z-Image Loras i've been training. Prior seeds will change. I've finally added a certain feel and style to the Lora.

v2.5:

Unlike v2 this model actually might need to be increased in strength. Try 1.2. It's also 3gb due to reasons.. Working on a fix in the mean time.

v2:

Might be a bit too strong, mistakes were made.. Try 0.4-0.7

ABOUT THIS LORA:

This is a Z-Image low rank 4-bit LoRa model trained on images from my UwU_Clarity SDXL model. This is meant to be a fast lightweight LoRa with low impact on VRAM. It aims to bring my SDXL models style and tone into Z-Image. This includes anatomy, genitalia, characters and concepts. Z-Image is great with realism but lacks the ability to produce fantastical images without a bit of creative prompting, this LoRa aims to address that making it easier to prompt for results like Flux and SDXL. Additionally out of the box Z-Image struggles with sexual concepts and censors the male member AGGRESSIVELY, resulting in horrific nude male figures. I've to some extent gotten the LoRa to a point that alleviates some of monstrosities the base model creates. This LoRa is a GENERALIST model, and doesn't specifically specialize or accel in one task.

QUIRKS:

Much like my base model, this LoRa heavily favors creating men and women that look distinctly unique/same-y. Due to my preferences of men and women. So to break away from people/characters that look the same I'd advise creating unique names + race/nationality + distinctive features when prompting.

(EX: Alexa Valencia a pale swedish woman with a wide sleek nose, lush lips, slender face and small serious eyes.)

DISCLAIMER:

This LoRa is NOT perfect! You will likely still encounter strangeness and default behaviors native to Z-Image Turbo. This is still VERY much a work in progress until I can figure out the quirks of the text encoder and the model itself.

TRAINING INFO: UPDATED!

The dataset like I mentioned before is a mix of my UwU dataset and a tailored synthetic dataset that I created with Z-Image and my UwU_Clarity SDXL model. I created a series of images with my SDXL model, encoded the resulting image to do a img2img pass with Z-Image. The workflow uses promptgen to caption the image created by SDXL and appended the original prompt used to create the image with promptgen's prompting style. Creating a caption method that supports natural language and a traditional tagging styles. (I can provide the workflow if needed. Workflow utilizes a MultiGPU setup but can be ran on a single card.)

The LoRa was trained to roughly 40k steps using 5k images containing a mixture of art styles varying in realism (anime, semi-real, realistic, photographic). This prevents leaning into a single style but also created a drop to the base models generations for photo realism. If you find the LoRa to be too strong lower the weight gradually until you're happy with the results. (Adversely lowering the strength will cause genitalia to revert to distorted lumpy meat sticks.)

I've now additionally merged in my other z-image loras in a attempt to further enhance the overall generations. Each were trained with different settings and the merges in various strengths.

Trained on a RTX 3080 12GB VRAM + NVIDIA QUADRO P1000 4GB + 64GB DDR4 RAM. The P1000 runs the training UI + Python and does some small caching vs using strictly CPU offloading + Caching. This makes it possible to use the full 12GB of VRAM on my RTX 3080 for training. This offers a small speed-up in training without OOM Errors and stops.

ADDITIONAL TRIGGER WORDS: (Unique Tags From My UwU_XL_v3 Model)

pony_style, illustrious_style, illustrious, pony_sdxl, semi-real, photoreal, uwu, lustful_clarity, supremely detailed, insanely detailed, interesting composition, digital illustration, animated image, absolute realism, highres, incredible shot, No description generated, No tags generated, Featuring stylized artwork, newest, high quality, wonderful, masterpiece, best, top rated, perfect image, feministic, ideal beauty, absolute realism, professionally drawn, theatrical color grade, liminal space, isolated spaces, dramatically eerie, dread and despair, minimal setting, absurdres, z-image style, penis, phalic, cock, dick, pubicle, gentials.

NEGATIVES:

(fleshy growth, ball of flesh, tumor, skin mass, lumpy genitals, abscesses, lipomas, inflammation, cyst, lumps and masses, disfigurations, skin condition, abnormal skin, 3 legs, three legs, tri-limb, elongated, saggy skin, ribbed pattern, diamond pattern, raised grooves, waffle pattern, finned, flanged, beef stick, beef jerky, meat stick, sausage, warts, pimples, ugly, disgusting, monster, creature, horrific, stretched body, stretched limbs.) + (your normal negatives)