IMPORTANT:
The default model that's available for download is overtrained and will steer heavily toward the default appearance. The prompt adherence to words that change that is very low. To get a more versatile checkpoint, please use one of the other checkpoints that are included in the attached zip file - the lower you go, the more malleable it will get (likeness will also reduce by a bit).
Will rework the training workflow and try to get a better model out soon.
Flux Retrained
Retrained model with the same dataset. I have uploaded the 4000th epoch but some of the earlier epochs might be better. Generating a single image takes >2 minutes on my device so haven't been able to do a good comparison.
Depending on the feedback on this version, will test and release earlier step.
NOTE: I have added multiple checkpoints in a zip file to the model. The default version will have very strong likeness but if you'd like to test earlier steps for better variance, feel free to find a step you like. If you manage to get better gens out of any of the other checkpoints, please post them here and share your params in the comments.
Flux
Recommended strength is 1.
This version was trained (21, 25 images) on Flux.1 Dev. Picked the last model at 4000 steps. Trained on an A40 on runpod - took 3 hours.
I tried running it on schnell but it didn't work, if somebody can figure out how to run it on schnell/schnell-dev-merge, that'd be great - reach out.
Pony
This version was trained (18, 23 images) - (24, 16 repeats) on Pony Realism. Picked the 6th epoch.
The face likeness is pretty good IMO, further training did not improve it.
I have included an image where I used facedetailer/inpainting for the face with my SD 1.5 model. This provides the best output IMO.
Also added a comparison of Pony facedetailer vs 1.5.
V2
LORA trained on 30 publicly available images of Emma Watson.
This model comes with an optional trigger word h3rm10n3to replicate Hermione Granger from the Harry Potter movies.
Trained on base SD 1.5, should work with most models.
Use ADetailer if details are lost and strength of 0.8 should work on most prompts.
Consider leaving a review if you like the model. All feedback is appreciated.
V1
LORA trained on 90 publicly available images of Emma Watson.
Trained on base SD 1.5, should work with most realistic models.
Use ADetailer if details are lost and strength of 1 should work on most prompts.
Description
FAQ
Comments (43)
In your opinion which model performs better and is mor elike her?
Is flux superior to alternatives or not?
flux is pretty new and the support for a lot of things is still being built.
Each version has it's purpose
SD1.5 - fast generations, lot of tools/extensions, extremely good likeness
Pony - likeness is hit or miss, but prompt adherence within SD is great BUT
Flux - likeness is really good and prompt adherence is excellent as well
I am not an expert at generating the best images but SD1.5 has the best community tools as of right now.
If development keeps going at the same rate, flux might get there far sooner than anybody expected.
If your system resources are limited, SD1.5 is still the best bet.
Generating a single image on flux dev takes me longer than a minute.
One thing I've noticed about Flux is it doesn't burn out easily.
For example, in an SD1.5 LoRA - the likeness starts in the 4th epoch and reaches convergence at 7 but burns out after.
in Flux, it reached likeness at step 750 and even at step 4000, it wasn't overtrained.
TLDR; SD1.5>Flux>>>>>>>Pony
@__2_ thank you. WIll try to install flux on my comfyUI
Just me or this doesn't show up for Pony?
not sure what you mean.
are you using the right model?
the pony model was trained on Pony Realism - will work best on that checkpoint.
haven't tried it on others.
@__2_ Not sure, it's just not populating for me in Pony or Pony Realism. I'll keep trying.
Turns out there was a bug in training with the tool I was using GitHub - ostris/ai-toolkit: Various AI scripts. Mostly Stable Diffusion stuff.
Fixed now, but it requires a full retrain. Crap!
the current model will work but the new one might have better fine grained detail.
not sure if it's worth a full retrain though.
@__2_ definitely worth it. Grainy dots are terrible.
Released the retrained version. @guy33 try the new model and let me know if the grainy dots are gone.
@__2_ I tried your first version and it was definitely inferior to the other model on Civit. Very fuzzy end results, though, good flexible likeness. I will download the new version and try it out in a while. The other Emma model has great likeness and clean image quality.
@SuoixDs good to know. Since this model takes a while to run, I haven't been able to set up a good benchmarking strategy to compare different step checkpoints.
Can you share some prompts/signs to look for when using flux loras?
Also, can you share the model/settings you are using for generations?
@__2_ I posted 5 images, but removed one that it was marking as pg-13. I had 5 images of mine removed yesterday because I did not realize you could not have non-sexual shirtless images depicting famous men. Anyways the prompt I use is there. I generate 1024x2160 images in about 37 seconds. The images have no adtailer or hires work done on them. I find that 1024x2160 works really well with FLUX for doing native image generation with out trying to put the subject in more than once.
@__2_ I guess my concept for what to look for depends on how much flexibility you want to have with it. On the new Arnold Person model, I can have him dress in Cosplay while still looking like Arnold or I can have it look like Arnold if he had feminine features or was maybe the sister of Arnold. Where there is a strong resemblance but not exactly him or exactly him. I did a ton of weird ones with Arnold. On your LORA, though, Emma almost always looks like Emma, which is a good thing, but it doesn't seem to want to follow prompting to pull away from her default appearance. This could potentially be resolved on the user's end by decreasing the strength of the model, but I do all the default testing on a default setting of 1. Personally I want to be able to make Emma into a space alien if I want like as if she was from Star Wars or Star Trek or some varied role playing race. But I also want it to be cinematic quality. lol No hard request right? :) Don't want to discourage you. I just have had much better overall results with the one made by Punzel. If I had not used the other, I would still use this because the likeness is great. It's just not as crisp and flexible. The crispness might just be because of source images used, but I don't know. If someone generates lower size images and does hi-res it might not even matter.
@SuoixDs Thanks for the context - feedback is always welcome.
I looked at the other model as well and it's mentioned that it was trained for 1600 steps. This model is 4000 steps deep. With SD1.5 loras I always select the latest epoch that hasn't burned (to get the best likeness) - that might not be the way to select models for Flux. I will try your prompts on earlier steps and see if they perform any better.
And regarding the crispness, that might be my fault, I used 512x512 images for training and all my testing generations were at a similar resolution so they looked fine.
Will try to get a better workflow going for the next model.
@__2_ I have a really good set of 20 example Bad Ass Heroine images I did for punzel's model, but for whatever reason, it is not showing up even though I posted it hours ago. All of the images I thought were decent/respectable, so not certain what the deal is. Hopefully they show up later, as they would be a really nice comparison for you I think since they use the same prompt except for trigger word. Just realized it flagged the entire set of images for review even though they all showed up as PG... No idea. lol
@SuoixDs are you using dev fp8?
i am trying to replicate your images on my comfy. getting different results on flux dev.
@__2_
unet:flux1-dev-fp8
weight: fp8_e4m3fn
clip 1: clip_l
clip 2: t5xxl_fp8_e4m3fn
strength_model: 1
strength_clip: 1
vae: ae
FluxGuidance: 3
denoise: 1
Those are the other settings I have on my workflow. Oh, and the images would probably look just fine at 512, but I test out what each LORA is able to do before upscale since I have a 4090.
@SuoixDs https://civitai.com/posts/5470331
For some reason the lora doesn't even work on fp-8 for me on runpod comfyui. so weird.
But generated some samples with your prompt and settings with earlier checkpoints and regular dev.
Check it out.
@__2_ definitely adheres to the prompt much better now. One has a really funny head size, and some of the hands are definitely janky, but at that point it lets you guide it in the direction you want while maintaining the likeness.
@SuoixDs I have added quite a few more earlier checkpoints to the model in a zip file. It'd be great if you could it test and compare against the prompts you liked on the other model and see if any of the checkpoints come close.
@__2_ I have a busy day, so I will try to get to it later. I will download it for now.
@__2_ Real quick, I just saw my other post was finally reviewed and approved. If you want you can quick see how it handled the Bad Ass Hero test. https://civitai.com/posts/5463022
eeeh its like.. 70% of the way there. Needs a little more time in the oven mio.
Can you check this post?
Emma Watson flux retrain progress | Civitai
why are these images that were generated as samples during training so much better than the ones i generate locally?
Is it because they were not generated with fp-8 version?
These steps performed so well during training sampling but on local they aren't even close.
Here are some gens on regular flux dev with earlier checkpoints - https://civitai.com/posts/5470331
is this a known issue with flux loras? (the disparity between fp-8 and regular gen)
@__2_ maybe you should training it with SimpleTuner?
@OrangeJuiceAlien I could but punzel has already released some really good models showing that that approach works. Just trying a different approach for now.
Not quite Emma Watson. As someone else said, it's about 70% of the way there.
Did you try the model on local? on which checkpoint?
I seem to be getting better results in forge with the same lora and dev-fp8.
Can you post some samples?
the posts by other users posted below look pretty close too.
my generations in the gallery above are comparably worse. (don't just* go by those - will replace them with better examples soon)
In my experience Emma Watson loras tend to require higher guidance than others. It only started to work for me without a doll-like appearance between 4 and 10 (with a lora strength of 1.0), and that's without prompting her name (which pulls in the awful version of her that's in all the base models). I can get other character loras down as low as 2.0 before they start breaking.
does that work on dev nf4 model?
yes, tested on forge.
@__2_ hey so i just have to use it in <lora:1> format or another way? cause I tried but the images were nowhere near emma
update forge to the latest
use this
flux1-dev-bnb-nf4-v2.safetensors · lllyasviel/flux1-dev-bnb-nf4 at main (huggingface.co)
a picture of 3mm4w <lora:3mm4w-flux:1>
this should work
NOTE: when running this lora with nf4, my 3080 10GB runs out of memory and has to retry
but it works
Could you please share the training config file full settings? I would like to use it as a template. Many thanks in advance!
Is the Flux Lora trained at 512x512 pixels? It is pretty terrible quality.
Just made a pics of (almost) Emma Watson in Flux Schnell.
out of all the Emma loras I tried so far, this one worked the best for me.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.