Another experiment that came out well, the idea was simply to make a lora that can do alot of the post processing i end doing to my images like adding film grain, edge blur, muted "cinimatic"/moody color grading etc.
I think along with ZIT's great prompting this can make pretty good images. Made it for personal use but i think some people might find a use for it :D
I think it can start to make strange artifacts with some schedulers/nodes, i stuck with the eular/simple combo to make it easy for people to replicate at 1.0 strength
PS. hands are a little broken i think no idea what i did to do that :(
Feel free to let me know what you think.
Description
FAQ
Comments (28)
Epic lora mate, I'm getting beautiful images🔥
had a feeling you will like it :D Thank you
@purplelady The grainy and contrast effect is great if tuned correctly with the right prompt; I'm going to use it from now on, hehe.
What a wonderful LORA, make picture very dramatic and moody, super~!
0.8 strength is work perfect for me.
Thank you :D I really like what people have been doing with it too :D
Amazing LORA! Precise, creative and so full of style. Thank you so much, purplelady!
thank you :D
as always i do this stuff to make my own stuff, but i am loving what people can do with it :D
Any plans to make ZIB version of this nice LoRA? I really like how it works, but ZIT output feels so limited after using ZIB for a while.
i actually havn't tried ZIB yet, i downloaded it the day it game out and it has just sat on my computer.
but yeah eventually ill retrain it, good thing is retained the dataset, i guess ill give it another try over the weekend
@purplelady Would love to see a ZiB version of all your ZiT loras. I'm still using ZiT for generation, but loras trained on ZiB work sooo much better (ZiB loras are ZiT compatible)
@nonamezoname7621 so that is kinda the plan, I have all the lora datasets setup with some improvements etc. once Z-image base training is available on civitai i am going to start publishing those.
Thats interesting when you say loras work better for ZIB i havn't messed around with it too much but i guess that might be one thing to properly try :D
And thank you, i actually have 2-3 more loras pretty much trained and ready just been lazy in making the showcase images to publish them :)
Bro, you can use this Lora with Z Image Base without any problems, but you should adjust the thickness. I'm currently using it on my own ZIB model with a strength of 0.60.
==================================================
LORA OPTIMIZER - ANALYSIS REPORT
==================================================
Architecture preset: dit (DiT (Flux/WAN/Z-Image/LTX/HunyuanVideo))
Detected architecture: Z-Image Turbo (Lumina2)
Key normalization: enabled
--- Per-LoRA Analysis ---
ZImage\Z-Image-Fun-Lora-Distill-2603_UDCAI_ComfyUI.safetensors:
Strength: 0.4
Keys: 180
Avg rank: 64
L2 norm (mean): 6.3508
ZImage\zitMj02.safetensors:
Strength: 0.5
Keys: 240
Avg rank: 32
L2 norm (mean): 0.6137
ZImage\Purple_grainy_zit.safetensors:
Strength: 0.6
Keys: 240
Avg rank: 32
L2 norm (mean): 3.6544
--- Auto-Strength Adjustment ---
ZImage\Z-Image-Fun-Lora-Distill-2603_UDCAI_ComfyUI.safetensors: model 0.4 -> 0.3400, clip 0.4 -> 0.4000
ZImage\zitMj02.safetensors: model 0.5 -> 0.4250, clip 0.5 -> 0.5000
ZImage\Purple_grainy_zit.safetensors: model 0.6 -> 0.5100, clip 0.6 -> 0.6000
Model: scale factor 0.8500
Model: orthogonal floor 0.85 applied for zimage to preserve independent contributions
Model: exact streamed energy 116.4399 (orthogonal baseline 116.4388, avg cos 0.000 — mostly orthogonal (independent))
--- Pairwise Analysis ---
ZImage\Z-Image-Fun-Lora-Distill-2603_UDCAI_ComfyUI.safetensors vs ZImage\zitMj02.safetensors:
Overlapping positions: 17995834
Sign conflicts: 8997582 (50.0%)
Magnitude-weighted conflict: 50.0%
Excess conflict over cosine baseline: 0.1% (expected 50.0%)
Cosine similarity: -0.001
Subspace overlap: 0.01
ZImage\Z-Image-Fun-Lora-Distill-2603_UDCAI_ComfyUI.safetensors vs ZImage\Purple_grainy_zit.safetensors:
Overlapping positions: 17997871
Sign conflicts: 8998199 (50.0%)
Magnitude-weighted conflict: 50.0%
Excess conflict over cosine baseline: 0.1% (expected 50.0%)
Cosine similarity: -0.000
Subspace overlap: 0.01
ZImage\zitMj02.safetensors vs ZImage\Purple_grainy_zit.safetensors:
Overlapping positions: 23970137
Sign conflicts: 11941828 (49.8%)
Magnitude-weighted conflict: 48.9%
Excess conflict over cosine baseline: 0.0% (expected 49.4%)
Cosine similarity: 0.017
Subspace overlap: 0.03
--- Collection Statistics ---
Total LoRAs: 3
Total target groups: 240
Avg sign conflict ratio: 49.9%
Avg weighted conflict ratio: 49.7%
Avg excess conflict: 0.1%
Avg subspace overlap: 0.01
Importance ratio (max/min frobenius): 6.71x
Decision smoothing: 0.25
--- Auto-Selected Parameters ---
Merge mode: weighted_average
(global fallback — each target group uses its own parameters)
Strategy set: basic (TIES vs weighted_average only)
--- Per-Group Strategy ---
weighted_average (orthogonal): 240 groups (100%)
Total: 240 groups
--- Block Strategy Map ---
layers.0 ---- avg 0% conflict (8 avg) (8x)
layers.1 ---- avg 0% conflict (8 avg) (8x)
layers.10 ---- avg 0% conflict (8 avg) (8x)
layers.11 ---- avg 0% conflict (8 avg) (8x)
layers.12 ---- avg 0% conflict (8 avg) (8x)
layers.13 ---- avg 0% conflict (8 avg) (8x)
layers.14 ---- avg 0% conflict (8 avg) (8x)
layers.15 ---- avg 0% conflict (8 avg) (8x)
layers.16 ---- avg 0% conflict (8 avg) (8x)
layers.17 ---- avg 0% conflict (8 avg) (8x)
layers.18 ---- avg 0% conflict (8 avg) (8x)
layers.19 ---- avg 0% conflict (8 avg) (8x)
layers.2 ---- avg 0% conflict (8 avg) (8x)
layers.20 ---- avg 0% conflict (8 avg) (8x)
layers.21 ---- avg 0% conflict (8 avg) (8x)
layers.22 ---- avg 0% conflict (8 avg) (8x)
layers.23 ---- avg 0% conflict (8 avg) (8x)
layers.24 ---- avg 0% conflict (8 avg) (8x)
layers.25 ---- avg 0% conflict (8 avg) (8x)
layers.26 ---- avg 0% conflict (8 avg) (8x)
layers.27 ---- avg 0% conflict (8 avg) (8x)
layers.28 ---- avg 0% conflict (8 avg) (8x)
layers.29 ---- avg 0% conflict (8 avg) (8x)
layers.3 ---- avg 0% conflict (8 avg) (8x)
layers.4 ---- avg 0% conflict (8 avg) (8x)
layers.5 ---- avg 0% conflict (8 avg) (8x)
layers.6 ---- avg 0% conflict (8 avg) (8x)
layers.7 ---- avg 0% conflict (8 avg) (8x)
layers.8 ---- avg 0% conflict (8 avg) (8x)
layers.9 ---- avg 0% conflict (8 avg) (8x)
Legend: ==== sum ~~~~ slerp ---- avg ++++ cons #### TIES
--- Reasoning ---
Excess conflict 0.0% <= 25% threshold -> weighted_average mode selected
Low conflict means LoRAs are mostly compatible, simple averaging works well
--- Merge Summary ---
Keys processed: 240
Model patches: 240
CLIP patches: 0
Skipped keys: 60 (shape mismatch, e.g. sliced weights)
Output strength: 1.0
CLIP strength: 1.0
Suggested max output_strength: 1.00
(energy preserved — no compensation needed)
==================================================
hello! is there any comfyui workflow?
they should be embedded in my images. however i am not doing anything really different from the default workflow you will find from comfyui etc. for these images
the obvious sticking effect is present
多几次尝试就会有惊喜，最喜欢的lora，谢谢
Thank you
Any plans for Klein or x base?
i can try Klein, for Zimage base civitai will need to actually add that to training models first :D
@purplelady awesome, please 9b Klein lady :)
I gave it a try and it came out really weird in 9B, ill change some stuff and try again, but it looks like Klein really wants to not having filmgrain and stuff as part of the image.
@purplelady damn, that's some bad news
@Neon_signs yeah ill give it another try with a slower LR then default maybe it will help, i had luck with it when doing zimage loras before
@purplelady awesome!