Update: Please if you see any of my models somewhere else report or let me know! Thnx
Trained on base model xl 1.0 + F1D + zit + Illu + Pony + SD1.5
Please read about version info (about this version box)
Feel free to check Doomguy: Status bar "bloody facial expressions as well
Feedback and ratings are much appreciated.
Feel free to check other similar models:
Shell Shock facial expression "Combat stress reaction" style XL
Crying sad (tears) style XL + SD1.5
Rotated View (Dutch Angle) Cinematic Style XL
Cinematic Split Screen Style XL
Classic + (Neo) Film Noir Style XL
Fine Art Creative (Cinematic) Photography style XL
Ron Fricke Film Director and Cinematographer style XL
Kodak Film Grain (Cinematic) Photography style XL
Cinematic Photography Style XL
Cinematic Explosion Style XL + SD 1.5
Cinematic Kodak Motion Picture "Film Still" Style XL
Realistic Photographic Film Stock Footage Style XL
Cinematic Film Stock Footage Style XL
Cinematic Vintage Film Style XL
Description
F1D v2
Lower lora weight when needed.
Should help with facial expressions using flux.
Trigger words not necessary but description is essential. (Happy, sad, scared, etc.)
I suggest you use Crying sad (tears) style for better teary eyes with this model.
PS. If you don't like this version you can always go back to previous one.
FAQ
Comments (7)
When I run my SAFETRIGGERS tool, your LoRA appears to be trained for the following TAGS, SORTED IN ORDER OF PRIORITY.
https://civitai.com/models/1343150/safetriggers-python-tool-for-reading-tags-and-triggers-from-loras
I hope this helps.
[["sharp", 920], ["solo", 806], ["skin pore", 801], ["movie themed", 800], ["detailed skin", 800], ["epic cinematic photography", 800], ["artistic", 800], ["creative", 800], ["wrinkly face", 800], ["dramatic lighting style", 800], ["cinematic color style", 800], ["kodak film style", 800], ["emotional style", 800], ["gustier style", 800], ["facial expression style", 800], ["realistic", 724], ["looking at viewer", 643], ["1girl", 575], ["lips", 501], ["photorealistic", 481], ["portrait", 464], ["brown hair", 428], ["1boy", 409], ["teeth", 403], ["short hair", 403], ["open mouth", 391], ["black hair", 390], ["male focus", 389], ["simple background", 350], ["smile", 348], ["long hair", 326], ["brown eyes", 283], ["closed mouth", 264], ["shirt", 264], ["upper body", 260], ["blue eyes", 229], ["blurry", 213], ["nose", 206], ["close-up", 187], ["blonde hair", 187], ["breasts", 185], ["closed eyes", 176], ["freckles", 168], ["facial hair", 165], ["white background", 161], ["nude", 161], ["black eyes", 157], ["sad", 153], ["grin", 151], ["blurry background", 145], ["jewelry", 135],... and so on!
...
Note. "sharp" : 920, "emotional style" : 800, "facial expression style": 800
So 120 images aren't tagged with references to expressions, the purpose of the LoRA?
Might this explain user's inability to get it to activate?!
Hope this helps.
This "tool" has nothing to do with model training! lol
The model has been trained on many facial expressions. (Happy, sad, scared, angry, etc)
I keep updating model(s) with better version(s) whenever I have some time.
You can lower lora weight for better results.
You can try different version(s) of the model. (For example, v2 might be better in some expressions than v3 and vice versa)
I believer (I can be wrong) Flux 1 Dev. use knowledge distillation or model distillation processes. Which can be the cause of dissatisfying in some cases.
Some argue an official release of a "De distilled Flux 1 Dev. model) which can help many creators better fine tune images with better image results.
Finally, I suggest, you ignore all this and focus on what exactly you are trying to generate. Think of this lora as an extra helping tool. For example, you want some one looking scared, you can try typing, words like (A scared man with big surprised eyes staring at camera in fear and worry and open mouth) And figure best weight that satisfies you! For example, you can go with 0.8 and get good results with some expressions, and you can go with 0.7, 0.9, etc. and get other good results with other expressions. Additionally, you can try mixing lora's, For example, Facial expressions model + Sad crying model + Shell shock expression model. etc. (Might help more)
Hope this helps with your case.
@EauDeNoire I wrote it. I haven't used your LoRA. Perhaps looking more closely at your tags in your images may help?
@Davros666 Any lora need to be tested, some may find it useful, others may find it useless. It's a matter of taste, I guess... These tag words you mentioned are small part of many other tagged words used in the lora model. (That tool only gives you what it sees not how lora actually has been trained!)
Image showcase might help a bit, still, any lora, needs to be tested.
In addition, how many emotions can a face really make?! lol
@EauDeNoire "others may find it useless" Ever wonder why?
@EauDeNoire "That tool only gives you..." Tell me what you understand about how you think tags in tagged images work...
@EauDeNoire "In addition, how many emotions can a face really make?! lol"
Humans can make thousands of facial expressions (over 10,000!) using 43 muscles, but scientists often categorize these into core sets like Ekman's 7 universal emotions, newer studies identifying around 16 common emotional expressions, or even more nuanced models with 21 distinct emotion categories and their combinations
Annnd blocked..
Not to mention nearly everyone of your models, most people can't even replicate anything remotely trained as represented in your images. Usually a big red flag that someone does not know what they are doing, is when the rest of the community can't even inject anything tied to your examples.