LoRA extractions is not as good as the full model but it creates a compromise between size and quality and it can be used along other LoRA model with a full model. Credit goes the creator of the checkpoint in version description
Sorry, i do not accept anymore request link(s) in my comments "Suggestions for a model extraction?"
04.0417.23 Name version is limited to 10 25 character, make it impossible to see full name... see below "List of all LoRA model".
How to use it in AUTOMATIC1111 WebUI?
Place the LoRA model in .\stable-diffusion-webui\models\Lora then use <lora:filename:multiplier> in the prompt.
About the comparison:
Picture order: First = LoRA, Second = Fine-tuned Model, Third = Base Model
DM = Network Dimension ( 120=lower quality/low size, 320=higher quality/higher size ) same for CD = Conv Dimension (Rank)
[CD is optionally if you wish to use it, install the extension: https://github.com/KohakuBlueleaf/a1111-sd-webui-locon (How? Extensions -> Install from URL)]
Obviously Version "1024CV1024" or in short "1024" use more VRAM
I did a large comparison DM+CV heresample pictures does not use vea and LoRA model with SDv1.4
every model from (Stable Diffusion Base Model: mostly sd-v1-4.ckpt, model base is determined by highest % affinity) was extracted using Kohya_ss (Save precision: fp16)
Name format: <Type><DM> <CV><Name>
List of all LoRA model:
LORA1024CV1024experience_80 - LORA320experience_80
LORA1024CV1024cardosAnimated_v20 - LORA320cardosAnimated_v20
LORA1024CV1024flat2DAnimerge_v10 - LORA320flat2DAnimerge_v10
LORA1024CV1024realismEngine_v10 - LORA320realismEngine_v10
LORA1024CV1024vectorArt_pulpVectorBeta - LORA320vectorArt_pulpVectorBeta
LORA1024CV1024hasdx_hasedsdx - LORA320hasdx_hasedsdx
LORA1024CV1024verisimilitude_V14 - LORA320verisimilitude_V14
LORA1024CV1024galaxytimemachinesForyou_v10 - LORA320galaxytimemachinesForyou_v10
LORA1024CV1024galaxytimemachinesGTM_v3Pruned - LORA320galaxytimemachinesGTM_v3Pruned
LORA1024CV1024hassakuv1m
LyCORIS1024CV1024illuminatiDiffusionV1_v11 - LORA320illuminatiDiffusionV1_v11
LORA1024CV1024xrikishisUnstable_alpha50 - LORA320xrikishisUnstable_alpha50
LORA1024CV1024diivesStyle_10 - LORA320diivesStyle_10
LORA1024CV1024elldrethsImagination_v10 - LORA320elldrethsImagination_v10
LORA1024CV1024digitalVividMemoriesAnime_10 - LORA320digitalVividMemoriesAnime_10
LORA1024CV1024generalPurpose - LORA320generalPurpose
LORA1024CV1024dreamlikeDiffusion10_10 - LORA320dreamlikeDiffusion10_10
LORA1024CV1024epicDiffusion11 - LORA320epicDiffusion11
LORA1024CV1024protogenInfinityX86 - LORA320protogenInfinityX86
LORA1024CV1024protogenNovaExperimentalX80 - LORA320protogenNovaExperimentalX80
LORA1024CV1024protogenX58RebuiltScifi - LORA320protogenX58RebuiltScifi
LORA1024CV1024modernDisney_v1 - LORA320modernDisney_v1
LORA1024CV1024revAnimated_v11 - LORA320revAnimated_v11
LORA1024CV1024revAnimated_v122 - LORA320revAnimated_v122
LORA1024CV1024comicBabes_v1 - LORA320comicBabes_v1
LORA320CD128perfectWorld_v2Baked - LORA120CD128perfectWorld_v2Baked -LORA1024CD1024perfectWorld_v2Baked
LORA14DM320realisticVisionV20_v20 - LORA14DM120realisticVisionV20_v20 - LORA1024CV1024dreamshaper_4NoVae
LORADM320realisticVisionV13_v13 - LORADM120realisticVisionV13_v13
LoRADM320dreamshaper_4NoVae - LoRADM120dreamshaper_4NoVae
LORA14DM320grapefruitv41 -LORA14DM120grapefruitv41
LORA1024CV1024berrygold11 - LORA14DM120BerryGold-1-1 - LORA14DM320BerryGold-1-1
LORA14DM120chilloutmix_NiPrunedFp32Fix - LORA14DM320chilloutmix_NiPrunedFp32Fix - LORA1024CV1024chilloutmix_NiPrunedFp32Fix
LORA14DM120deliberate_v2 - LORA14DM320deliberate_v2 - LORA1024CV1024deliberate_v2
LORA14DM120HSPU-1.1 - LORA14DM320HSPU-1.1 - LORA1024CV1024hspu_11
LORA14DM120HSPUCar1-1 - LORA14DM320HSPUCar1-1 - LORA1024CV1024hspuCarV10
LORA14DM120Maiden - LORA14DM320Maiden - LORA1024CV1024maiden
LORA1024CV1024yiffye18Furry4Kinky
LORA14DM320protogenX53Photorealism_10 - LORA14DM120protogenX53Photorealism_10
LORA14DM320anything-v4.5-vae-swapped - LORA14DM120anything-v4.5-vae-swapped - LORA1024CV1024anything-v4.5-vae-swapped
LORA14DM320robo-diffusion-v1 - LORA14DM120robo-diffusion-v1 - LORA1024CV1024robo-diffusion-v1
LORA14DM320Silicon29 - LORA14DM120Silicon29 - LORA1024CV1024Silicon29
Description
base: diivesStyle_10.ckpt [68df4953]
v2-1_768-ema-pruned.ckpt [4bdfc29c] - -0.27%
v1-5-pruned.ckpt [a9263745] - 69.54%
sd-v1-4.ckpt [7460a6fa] - 85.50%
FAQ
Comments (28)
Good job, keep going ! There are so many good but big/huge models LORAble.
The dosmix-320 is gone, maybe deleted.
About the last published LORA, can you add the rev-animated-320 ?
I... am speechless.
This is... I struggle to find the words for how gigantic this is for us creators. I shall generate a statue of you in 3200x3200 resolution for this (once I stop my GPU from exploding), thanks thanks thanks.
Can someone tell me what this does? Use simple words, please.
Simply put, these are LORAS, aka agglomerate of logic behind a specific theme (for example: a LORA on a specific actor will tell the AI "Draw me this scene, using this specific model (example: Marvel Style comics) and put this actor in it". This is not limited to entities, though: a LORA can also force the AI to create a specific facial expression, color the eyes in a way, give a specific light to the scene, etc.).
These are, IMO, especially important because they are LORAS of big models, meaning you can now force the AI to use these "big models" for specific directive to give context, while also using the original model (a different one) for the theme. Example: You want to make a Fantasy Elf scene being drawn in DC\Marvel Comics style? Now you pick the RPG Lora (so that the AI knows what an Elf is, or a silver elvish armor, or a magic bow), and then use the "DC\Marvel comics" model to draw it. Before, you could not do that so easily (it was possible, but it was much more inconsistent).
@Vrane thank you kind stranger, this does indeed sound very useful!
Someone took a bunch of images generated from checkpoint models, and created LORA models from those outputs. So they are not trained on the same dataset but trained to mimic the look of what the various checkpoints output, usually to worse quality.
Nice but depends a lot on DIM RANK used for LORA extarction and since that's up to the model creator it's a bit random luck factor, Better to train same trained DB images via LORA than extract, at least from my experiments.
I would love to see Galaxy Time Machine, RPGv4, and Verisimilitude.
@herbert_west Please provide your request link(s) in my comments "Suggestions for a model extraction?"
Why not just use the kohya_ss script?
@ateist I have no idea how to do that...I'm somewhat proficient with SD workflow/modding stuff, but ultimately fall short of what many people here can do. There's a point at which the guides are a bit beyond my skill and comprehension. I've trained a few embeddings, but that's about it. So, I figure if someone is taking requests, why not?
@ateist how to extract a model using kohya
@monkeyman007 @herbert_west
use the script kohya-ss/sd-scripts/blob/main/networks/extract_lora_from_models.py
call it "python extract_lora_from_models.py --model_org <base_model> --model_tuned <tuned model> ----save_to <name_of_lora_file> --dim <size of LoRA> --save_precision fp16"
@ateist I have the only option to use colab. is there a way to do it on any of kohya's colabs?
@monkeyman007 it doesn't need any processing power. You can do it on any PC.
@ateist you would be surprised lol. all i have is 4gb ram thats shared with my igpu. also can i run the kohya script in automatic gui?
console is saying : Couldn't find Lora with name LyCORIS1024CV1024illuminatiDiffusionV1_v11
plus i'm pretty sure its not doing anything in the prompt maybe i'm doing something wrong , would appreciate the help
@watdafawx In prompt change name as it could be "CommunityLoraExtract_lycoris1024cv1024illu", don't forget to use locon extension
Do CarDos please
@DarkAgent Please provide your request link(s) in my comments "Suggestions for a model extraction?"
Why are there 2 files for each?
What is the difference?
@DarkAgent quality/size difference, more info in the description.
meinamix cannot be downloaded
@FlamingoDelta Fixed
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.