FLUX 🌟FaeTastic Details
Behind the scenes, we're working hard to get Flux onsite training ready here on Civitai! 😮 This test training aimed to see if a Detail concept LoRA could work on Flux—and it does! However, I do think the effect is less pretty compared to previous versions.
Regarding onsite training, we're hoping (fingers crossed) to release it maybe sometime this week... and if not this week, then hopefully next week. ⌚
Workflow and metadata comfy stuff should be in the images. I recommend trying out the workflow used here, as I can generate with Dev locally using a 4090 in 20 seconds. Additionally, I've come across other workflows where the LoRA actually has no effect on the image... so please ensure your workflow with LoRAs actually works. 😅
This was trained on Flux Dev. It does have some effect on Schnell, but the effect is not as strong or as impressive.
Description
FAQ
Comments (68)
I'm so happy! I've been hoping for this from you since flux released!
awesome, thank you!
This looks great! And good point about some workflows not having any effect on some LoRAs. Noticed it depends a lot on which workflow you use as well.
Yeah...someone was showing off their Flux LoRA in discord...but when I put their workflow in, the LoRA part of it didn't work period and it wasn't actually changing the image. Was quite funny 😅
@Faeia They probably connected the lora node incorrectly.
This is awesome! Thank you 🩷
Great work Fae!
Exciting news Fae! Thank you so much! :D
Love the style, super hyped for Flux training! 🚀
I can't wait to be able to do Flux Lora training. Thanks
Sad it doesn't work on NF4, or did somebody get it to work? 🙏
Sooooooooooooooo if i download it and use it inside of A1111 and generate iamges did i able to generaste any image with text effect now ??
As far as I am aware, Flux support is not available in A1111. Additionally, Flux base just generates text fine...the LoRA doesn't magically make text work...that's all the base model. :)
Flux support is available in Forge but without LoRa suport https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
@Faeia I found out today that dev also does text but only as 'speech bubble'.
Wow pretty amazing! Like its told "those little things in life.."
OMG! those sample images ... wow, I'm in love with all this artistic overload ;)
Does it also work with distilled models?
yes, using it with nf4 v2 in Forge: https://civitai.com/posts/5605200 and https://civitai.com/posts/5602817
I noticed when you change the Lora strength it changes the ear length of the geisha portrait. kinda weird lol. But other then that you can make some pretty crazy stuff
Hello,
I am here after reading the quick start guide to Flux, particularly its training section. There is mention of a different prompting system for training compared to SD. Did you experience the same as the article reports? If so, would you be willing to share all or parts of your training data? I would personally be delighted to see examples before attempting training on CivitAI. I only ask you as you are part of the team.
Thanks for reading.
What's the article?
So, I've trained with tag captions, natural language, and no captions. So far, no captions or natural language have been the best for me. However, I have only trained style and detail LoRAs. So things might differ for characters or concepts.
This LoRA was trained with Xflux...I believe it was 5 epochs and 5000 steps...and it was trained on both tags and natural language captions. This was one of the first LoRAs that I tested. When I get some time this weekend I'm going to try to train it again with Kohya a few times testing out some different settings to see if I can make the effect stronger like the XL version.
The Kohya LoRAs that I have trained so far for Flux, which I haven't released yet but will here soon once I get time to generate the preview images...I did a style one four times...
Tag Captions (Decent)
Tag Captions (Adding in more images to the dataset that I generated with the first one, including some realistic images) - This one went the worst
Natural language Captions with GPT4o (+ Removed realistic images, became better)...I need to test both Florence and Joy. Joy particularly is impressive to me. https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha
No Captions (Best one so far, and the one that I'm going to release)
The settings for Kohya was 10 epochs, 12 repeats ~ hit around 1k steps...and every time I would find that epoch 7-8 were the best. The rest of the settings are the ones we currently default our Kohya Flux training to.
So anyway, those are my findings so far that I can think of quickly off the top of my head. :]
@Faeia Sorry, I cannot seem to find that article again. Anyways, it was one of Civitai announcements about generation and only mentioned flux training briefly explaining what you pointed in your results, that natural language worked better than tags.
Thank you for your answer and details, I've used your settings for my first flux training attempt. Hopefully it will turn out alright.
It's a character lora, I have no idea if my description style will work properly or not.
@IndustrialVectors This? https://education.civitai.com/quickstart-guide-to-flux-1/
I'm really enjoying working with this lora, thank you so much for making it ❤️
are there some requirements for this using this Lora? I cannot get a clear picture, every thing is blurred
Are you trying to use it on our generator or locally? And if you're doing it locally, what are you using to gen with it?
locally， comfyui，flux.dev,
@aiamonlylove in Forge it works perfectly, and Forge currently is faster than comfy with Flux *
Agree with aiamonlylove, this LoRA does not work in ComfjyUI with Flux1-Dev. Bypass it, set clip and model to 1.00, or even 100.0, and no effect whatsoever.
Update: I take it back... This LoRA does work. You have to do an Update All in the ComfyUI Manager. And now that it works, I can say it's really quite nice!
works on nf4?
yes, using it in Forge: https://civitai.com/posts/5605200 and https://civitai.com/posts/5602817
@OliviaRossi thanks
I tried to use your workflow from your image, but a node is not laoding and comfy manager does not recognize either. It's ODESamplerSelect node
With GGUF you really have to turn down the weight of this LoRA or it's a mess! 👌
Is there a trigger word I'm missing? Because I'm trying with and without your Lora and the images are identical. (I'm using Krita with the ComfyUI backend.)
There's no trigger, I'm guessing it isn't using the LoRA then because it does change the image and a lot of the time significantly.
@Faeia I want to believe you. I'm using the Lora the same way I use any other Lora tho, and they work.
Same issue here, Flux1-Dev locally using ComfuyUI. It just doesn't work, no matter how I set the LoRA, or bypass it for that matter. A workflow from someone whom it works for would be nice.
Update: And right after I posted this I realized, Duh, the example images have the workflow. Aside from the custom nodes, they didn't differ much from my workflow. So, when things don't work, do an Update All in the ComfyUI Manager.... voila! Everything now works.... and without a need for the custom nodes in the examples, which I deleted.
@modelmixing Did you try updating Comfy like Modean suggested?
@modean987 thx. I had the same problem and a full upgrade fixed it for me too.
@modean987 I'll look into that. Comfy installed for me when I added this plugin. Not too sure how to do that but I'll figure it out - thanks.
There's no word to use in the prompt to trigger it (caption)? Thank you.
Very good lora, one of the most efficient i ever used. Unfortunately using other loras makes it less efficient but you can manage with lora weighting.
Also the nature of what you generate can makes it less efficient too. I think it's related to the training data used.
I was a fan since the start from your work and here you never disappoint, thank you !
Hi,
I've used the workflow in comfyui from the image but the text is not rendering legibly. It's like FLUX doesn't know how to render text. Can you help please? Nothing modified, added or changed.
When I bypass the lora text works.
So, the lora seems to break text and won't generate the same images as you have uploaded.
I have the same problem, try to change the lora strenght_mode to 0.8 instead of 1 on the node LoraLoaderModelOnly
@nelutu from 10 months in the future, I thank you, bruh. I was going crazy until I read your suggestion and lowered the weight to 0.8. BTW I'm using Comfy and the 15.9GB checkpoint version.
very good lora
Plz i have pc
Xenon 32 ram .. vram RX 5700 XT
I can download and run ?!
COOL! Love your work!
Hi, Faeia! I ran into a problem when loading your workflow. “When loading the graph, the following node types were not found: ODESamplerSelect”. But I can't find this ODESamplerSelect node. I just got in touch with ComfyUI recently and I'm still in the learning stage. Please forgive me if this question is a bit stupid.🙏
There is a button for installing missing custom nodes on a workflow. I don't remember exactly where it is right now, but it does work. I'm pretty sure it's in the manager section (comfy UI manager is a must-have)
@Squidbush Thanks a lot for your reply! I've found some other more useful workflows and I'm still learning them. Thanks for your guidance!
#AwesomeSauce
Very good lora (++ when you're sister/wife/daughter wants to be a fae ! 😊😂)
adsff
Like this one, works best not alone imho but combined with a 2nd lora if you want max results
Schnell also has a great effect, it's really cool
It's strange, but anime often produces results when you ask for realism, which greatly deviate from the original request. Forge UI, flux1-dev-bnb-nf4-v2
Not sure of the reason why all the positive comments this has... it instantly ruins my generates.
Not usable in InvokeAI:
Unable to determine model type for FluxDFaeTasticDetails.safetensors
Lots of potential
จินตนาการดี
What's the license? for using this model? Can I post the generated images on my instagram without credits?
Nice Lora! Thank you ^^
Details
Files
FluxDFaeTasticDetails.safetensors
Mirrors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
models-fs.safetensors
FluxDFaeTasticDetails.safetensors
FluxDFaeTasticDetails.safetensors
040125-25fandetails.safetensors
FluxDFaeTasticDetails.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.