--Anima
Trained on anima3. Recommended strength: 0.6.
Works best with block filtering — I created a ComfyUI nodepack for this: ComfyUI-LoRA-Block-Filter. Examples generated alongside my other LoRA — nicegirls.
Experimental version. Results may vary depending on prompt and block weight settings.
--Flux2
I use it with res2m+beta57, guidance 4, and it works very well. The quality is similar to my Qwen Lenovo, but I noticed that Flux2 is more flexible and has better prompt adherence.
Recommend to use with strength 0.85-0.9
-- Z-image Turbo
Actually i'm not sure that my workflow is correct, i believe the result could be better
Recommend to use with strength 0.80-0.9
I used euler + sgm_uniform with 4 guidance
P.S: at this moment for ComfyUI u need this pack to load lora properly https://github.com/PGCRT/CRT-Nodes
--Chroma
My first try to train lora that enhance realism in Chroma.
I'll upload my workflow today a lil bit later (or atm u can just take it from images) cause there some nuances about settings. But from my observations i noticed that for fast gen dpmpp2m + beta with cfg 3 works good . More experimental settigns are res_2s_rkmk2e + beta57 and cfg 3, more accurate gens, better limbs and fingers, but gives a bit weird grain on very small details + plus this method is pretty long
--Qwen
Works perfectly and much better then other versions. Gives extremely "live" images with amateurish quality with dramatic light and shadows, some blur and with vibe of 2000s street photography
You can take workflow here https://huggingface.co/Danrisi/Lenovo_Qwen/resolve/main/Qwen_danrisi.json
Thanx to fox23vang226. i understood that it's possible (and even have better effect) to use lora with values 1.5 - 2.0.
P.S.: almost all example images generate with both loras (NiceGirls)
--Wan
This is the first release — mainly a test to explore Wan's capabilities.
Note: In this initial version, facial details are somewhat lacking. For better results, I recommend using it alongside the Instagirls LoRA by my friend, which significantly enhances facial features.
--Flux
Who needs a fancy name when the shadows and highlights do all the talking? This experimental LoRA is the scrappy cousin of my Samsung one—same punchy light-and-shadow mojo, but trained on a chaotic mix of pics from my ancient phones (so no Samsung for now)
Description
FAQ
Comments (32)
Finally a F2K9B version of this awesomeness ! imo no model is worth using without your loras, even such awesome models as Flux 2 (both dev and klein) still cant compare to what they can do with your loras (Olympus and Lenovo) and it's not just the style they add, it's also knowledge of some concepts and stuff which original models have no clue about - thank you so much for making them !
The sinking feeling when I tried Danrisi's Flux2 models on Klein and it didn't work <<<
<3
Does lenovo_flux_klein9b.safetensors work with flux-2-klein-9b-fp8.safetensors? (not the base version).
Comfy gives lots of errors: lora key not loaded: single_transformer_blocks.0.attn.to_out.lora_A.default.weight. And resulting images are exactly the same as without lora.
Thx.
Hi. I guess u need to update comfyui
It's possible you're missing an update, myself as well, but I only see one error referring to a single block (using rgthrees powerloader).
If you're seeing application, or even better you can recreate an example seed photo, then it shouldn't matter.
@happylittleteapot thx. i updated comfy and it solved the issue. lora works now.
The klein-9b-base version is quite strong, even at 0.5 strength it changes half the anatomy (using 9b-base, trying it on 9b alone seemed to work closer to expected, but still strong). Will take more testing (I'm settling around lenovo:0.2, nicegirls:0.5). Nice_girls alone at 0.5 doesn't do nearly as much.
This is expected behaviour though (more extreme impact of loras on base), so nice.
Love seeing these get ported to Z-Image base. Thanks again
Just tried both the new ZiB and the Klein9b-Base loras. And no surprise here but the Klein9b-Base lora is the winner. This lora was the final touch I needed for some fantastic photos on K9b base. It creates a more natural skin tone while enhancing skin details. I don't think I need anymore convincing now that Klein is the future as it's way more trainable and it's flaws are being addressed by loras such as this one. Thank you for sharing this.
kelin 9b version is amazing, just what klein was missing for realism, I used it a1.5 .
Is it just me or is the OG Qwen w/ 4 step and other loras still the best model? Klein in particular seems to be way worse than OG Qwen.
just you
@goeypants400 Eh the picture posted below don't look better than the qwen lora, I don't get it
@greentheory yes because its low quality amateur photos, its masking it. same reason why videogame graphics look realistic if you put a vhs or body cam filter on it, but without it then the game goes back to looking non-photorealistic. doesnt matter if we get flux 3, the style of the low quality amateur photo was already achieved in flux 1, after that other things get improved which this flux 2 klein lora does so you are biased or giving only glances
I've been away from open source image generation for a while, I can't believe after all this time Civitai still did not solve the issue of LORA/models naming!? Why they don't have a structure! loraname-version-number-creator-name-model-name-date.safetensors
life is too short to waste time for good naming. sorry
@Danrisi I would hate for great LORAs and Models going to waste! I've a folder full of LORA names I don't even know which model those are for! that's a shame..
@mramer723 yeah, i completely understand you. but i have at least real name of lora in file. cause i still have tons of loras since pony and sd1.5 with names like "daskdji1h23912389hdasajd.safetensors"
@Danrisi at least its better than IDUNNOHOWTOCALLTHISLORA.safetensors lol!!
@meneper ahahaha. that was cringe
There should be a way for Comfy to know what LoRa goes with which model automatically (some kind of metadata tag), so at least you know what works w/ the loaded model. The naming system, I'm not sure that will ever be fixed. I just rename them, not too much effort.
"I used euler + sgm_uniform with 4 guidance"? How's that possible?
Why no? I think i uploaded this workflow to civit
@Danrisi There was no workflow in the two PNGs I tried. I don't feel like trying them all.
I thought Flux.Klein could only be used with CFG 1...
@CupRunethOverWithLSD it is for distilled, but not on base model.
but sgm and 4 guidance i wrote for zimage turbo. flux klein i use with 3 cfg and res2m + beta
@Danrisi Well, this is news to me that you can increase the CFG at ZIT so much without an additional guidance node. Will def try it.
First of all, thank you for your LoRAs. Some of the best on the site, as always.
How would you rank the models you've trained LoRAs so far?
I'm still very partial to Qwen, haven't used Flux 2 much however.
Does the qwen version work for qwen image edit models or just the base T2I model?
the realism is good but i cant do atractive pawg woman its always slim, small hips like there is some lock or something thats fuck me up man do something wit it
It's probably the training data. The model can only create what it knows.