27/05 : Updated with SVD INT4 Quant for Flux https://github.com/mit-han-lab/ComfyUI-nunchaku
Special thanks @theunlikely for the work done on quantification (took 6 hours on a H100 GPU), and thanks to JIB (J1B Creator Profile | Civitai) for his WF and clarifications on using this type of model:
For using this model you need to install the nunchaku project following the instructions there. and the nunchaku ComfyUI custom nodes to get it to work. Download and unzip the archive from Civitai to: \comfyui.git\app\models\diffusion_models\svdq-int4-flux-dev-de-distill - A Nunchaku workflow you can use : https://civarchive.com/models/617562
This is a repost from hugging faces, I did not play any role at all in the creation of this epic, groundbreaking model from nyanko7. I just had to post it asap because it is F A N T A S T I C
Flux.dev as we have known until now was a distilled model, meaning it was trained by flux.pro as its teacher. These new models change everything ! This is the first experimental, effectively de-distilled version of Flux.dev, meaning it is much closer to what flux.pro is capable of. And it's just the start !
(NB : This is not stating it was trained with flux.pro - I don't know the exact method)
/!\ READ ALL IMPORTANT STUFF BELOW (AFTER THE EXAMPLES) IN THIS DESCRIPTION OR IT WILL BE A FAIL FOR YOU /!\
EXAMPLES MADE MYSELF
Distilled CFG 8 VS Real CFG 8 - Fixed seed, no prompt change except for the text
Each picture is the first shot on each model : no cherrypicking / cheating
Ear gauge LoRa : "Cute blonde raver young girl smiling, facing viewer, with emo makeup and puffy emo hairstyle, green shiny colored hairstyle. 3argauge, both ears have Ear gauge plug with very large circular hole in her lobe. She is wearing a black hoodie with text "DEDISTILLED MAKES MY LORAS WORK" in golden letters
Cum on face Lora (with dedistilled it actually works anywhere) : "COF, Young woman with white sticky cum on face,white sticky semen on face,white sticky sperm on face,face covered with white sticky cum, face covered with white sticky semen, face covered with white sticky sperm. She is wearing a black hoodie with text letters "DISTILLED" in silver font on it"
No LoRa : "Dutch shot view of a black car speeding away from a massive explosion in the streets of a futuristic city with large buildings, towards the viewer, escaping the blast of the explosion. The motion lines around the car give the impression of speed. The front plate of the car has text "DISTILLED" on it."
No Lora : "A small kitten playing with a ball of yarn is seen through an old wooden window of a rustic house. The scene is cozy, with weathered wooden furniture inside the house and soft afternoon light streaming through the window, casting gentle shadows. Outside, in the distance, a photographer is approaching, camera in hand, ready to capture the playful moment of the kitten. The photographer is wearing a brown jacket and is framed by the soft glow of the golden hour light, adding a sense of warmth and tranquility to the scene. The overall atmosphere is peaceful, with a touch of nostalgia from the vintage setting."
SUMMARY OF THE IMPORTANT STUFF (aka current knowledge about it)
Disclaimer : These models are very new. So, just gathering here what is known about them for now. Please share your experiences in the comment section so we can update this together.
PARAMETERS
You can now forget about Distilled CFG and use real CFG (I've tried up to like 14)
NEVER, ever use it with CFG = 1 - It will automatically be a complete disaster, and most of the time the reason why you don't get results
You SHOULD use at least 40 - 60 steps, depending on the CFG you use.
It will be much longer but SO worth it
Unfortunately the current hyperdev 8-steps Lora doesn't seem to work with it to reduce steps
Dedistilled allows NEGATIVE PROMPTS
BENEFITS
Prompt adherence will be EXCEPTIONAL, even with Loras.
Faces Loras will work better, details will be better, text will be WAY better...
Everything from the prompt will be better basically
/!\ IF YOU DON'T SEE ANY IMPROVEMENT FROM DISTILLED MODELS : VERIFY YOU ARE NOT ACTUALLY USING REAL CFG = 1 WITHOUT KNOWING. NOT FLUX GUIDANCE. IT IS TRICKY /!\
AS ALL WORKFLOWS WERE TAILORED FOR DISTILLED, CONSIDER TRYING WITH FORGE IF IT'S NOT WORKING FOR YOU IN COMFY ?
GUIDELINES FOR USE IN FORGE UI
Works in Forge without any change. Will be loaded as if it was Schnell model, disabling Distilled CFG (cool).
EDIT : uploaded all the new Quants, you should find at least one working for you
If you are new to Forge, make sure you use similar settings :Flux workflow
DeDistilled as the checkpoint
In VAE / Text encoder files, provide the vae (ae.sft / ae.safetensors) + clip_l (or a modified clip) + t5xxl (whatever the quant you are using, fp16, fp8, etc).
Otherwise it won't work as they are NOT bundled in the model files herePlease also set Diffusion in Low Bits = Automatic (FP16 Lora), otherwise you might be in trouble with LoRas. This applies to any checkpoint in Forge.
I then recommend these settings for Dedistilled :
GUIDELINES FOR USE IN COMFY UI
Works in ComfyUI using a pretty standard workflow, the one cited below uses the GGUF Loader, Dual CLIP Loader for t5xxl and clip_l prompts, and KSampler Efficient
Recommended settings for Comfy :
Dual CLIP Loader guidance: 3.5 KSampler cfg: 2 to 10 Steps: 50 to 60 Negative Prompt: Can be left blank or can be provided if needed, will affect the image if provided
Sampler: DDIM or euler Scheduler: beta or exponential
WORFKLOW HERE : https://gist.github.com/dasilva333/87bdd5b5b8ebba5515a9919ede0e3c05
Found this one also on reddit (drag & drop it into Comfy) : https://files.catbox.moe/y99yl7.png
TRAINING AND LORAS
I have just trained myself my first LoRa using De-distilled and guidance = 6, after failing hard with Distilled and guidance = 1. Results are awesome, it basically saved my LoRa. Works great with both De-distilled and distilled (but better with De-distilled).
I will be using it to train from now on.
The first checkpoint fined tune with dedistilled has been posted on civitai here : https://civarchive.com/models/690991/sapianf-nude-men-and-women-for-flux-now-de-distilled
Awaiting answers from the author to update here
Sources
Dedistilled model FP16 : https://huggingface.co/nyanko7/flux-dev-de-distill
Dedistilled model FP8 : https://huggingface.co/MinusZoneAI/flux-dev-de-distill-fp8/tree/main
Dedistilled FP8 GGUG & Q4_ K_M GGUG : https://huggingface.co/TheYuriLover/flux-dev-de-distill-GGUF/tree/main
Description
FAQ
Comments (82)
When you say you can use the real CFG, I assume this means that Negative Prompt will innately work on this model? Since CFG:1 eliminates this function, being able to open the full CFG functionality will restore the use of negative prompts without workarounds, right?
Affirmative, although I just started using it yesterday and haven't tried negative prompts yet, but please update me on it so I can modify the description. Do you use Comfy or Forge ?
@valentinkognito365 I'm solely on ComfyUI, I'll see what I can do and let you know!
@Graybles cool, i have a friend that is struggling with comfy and it's clearly because his nodes are using CFG = 1. Let me know how it works with you and which nodes you are using :-)
First tests are good, Negative Prompt is being accepted in the standard ComfyUI workflow outfitted with GGUF loaders. It does indeed need at least 60 steps for the image to be perfectly sharp! You can leave the Negative Prompt empty and it'll still function, but Negative Prompt does work and can now successfully eliminate certain items or things from the render like it works in previous (SDXL and SD1.5) models!
Edit*: I went up to CFG: 8 without a problem.
@Graybles Great news :-) I have to try negative on forge. That would be great if you can make a side to side comparison with negative on / off, I just posted a Distilled FP8 vs Dedistilled Q8 GUGG one in the model pics. Also, can you tell me the names of the node you are using to input prompts and CFG and how they all called precisely ?
I remember ComfyUI uses stuff like Flux guidance, CFG etc. So i'd like to be sure what is actually distilled CFG vs Real CFG. Would help my friend and also people if I put it in the description here.
Hope you'll like playing with the models, they're slow but honestly if you're not using loras distilled is OK but as soon as you start playing with loras and complex prompts, I'd trade easily like 10 pictures made with distilled to have one made with dedistilled. It zeroshots prompts a lot !
@Graybles raising cfg higher than 1 effectively doubles my generation time - even with no negative prompt. Can you confirm this?
@2phazt Well, I made a Workflow for the distilled versions a bit ago and yes, using NegPropt (by using the "PerpNeg Guider" node) also lengthened the generation time to 2x time the generation time, I could just be that it's a condition requirement for using NegPrompt... As I couldn't find any other way to decently incorporate the NegPrompt without that node, not without losing quality. You can find the Workflow if you want to check it out on my CivitAI page.
@valentinkognito365 Did your friend figure out how to change cfg scale on comfy? I'm using Graybles gguf workflow and I can change the cfg scale there on the "Unified KSampler (Basic)" node, but not only the usual flux workflow nodes. Unfortunately I can't figure out how to link up Img2Img to that sampler node (the only one I can find that can change cfg scale for this model). The usual flux workflow nodes keep cfg at 1, resulting in bad images.
@eyu96934649 Here's the workflow I use https://gist.github.com/dasilva333/87bdd5b5b8ebba5515a9919ede0e3c05
It's the GGUF Loader + Dual Text Prompt Encoder + KSampler
What makes it interesting is that the prompt field has it's own cfg value but I haven't found it to have any effect on the image. I've also found that the cfg field in the KSampler behaves differently for destill vs non-destill models.
For the destill model I use a cfg range of 3 to 10, for non-destill models i use a cfg range of 1 to 1.8 in the ksampler, again the cfg field in the prompt node doesn't affect the image visually afaik.
@eyu96934649 If you are using my Current Workflow (probably v2.0a) it was make specifically for Distilled not this version of Flux. It's a Workflow that uses the Prependicular Negative Guider to activate Negative Prompt for Distilled Flux. But I don't believe it'll have a positive outcome on De-Distilled.
@Graybles @azimuthalobserver it would be great to upload your workflows on civitai at some point, we need to have as many people as possible use these models :-)
@valentinkognito365 The De-Destilled workflows are basically the Standard ComfyUI workflow but the checkpoint loader is a GGUF Loader, Dual Clip Loader and a VAE Loader. Nothing else, really...
Also, I do have the feeling that CFG is not cutting it on the detail of the image. It is in all the tests I ran worse in detailing than Distilled Flux as the Flux Guider you used to use is better at focusing on details in the image. I have decided to stay focused on the distilled version where I have the option to use Negative Prompts when I want. While it is true that the PerpNegGuider node lengthens the generation time by about 2x I have that same length all the time in De-Distilled and in De-Distilled I'm forced into 60 steps, that coupled with worse detailing is not really worth it for me!
@Graybles thanks for your testing and insights. I guess it's all depending on your uses cases and preferences. I really cannot go back to distilled except for easy stuff without Loras. I've just done a cum lora that was barely working on distilled and was immediately an almost overwhelming cum festival with dedistilled. Dedistilled allows me to make first-shot perfect generations of a lot of stuff I probably just couldn't do with distilled, or it would take a LOT of attempts. I absolutely love dedistilled and cannot wait for it to improve and become the standard for LoRas :-) It needs some hack to become faster though that's true. Can't complain, it's stuff that we would have considered a dream some time ago.
@Graybles I would be curious to see some of your comparison on details though, there's a few people saying the same but for me it's just the opposite
@valentinkognito365 I'll see what I can do in a couple of days, I don't have the De-Distilled model on my PC atm, but I'll redownload it. I'll see if I can create a decent Workflow for it as well, but this will probably not be before the weekend!
@Graybles I think you should persevere, seriously. I'm so amazed by dedistilled, I now think distilled is complete shit compared to it. It makes all my LoRas shine in a way that is straight impossible with distilled. I just made a new version of my POV blowjob LoRa that I will release in the hour. I challenge you to put the penis really deep inside the mouth with it and Distilled. It is very difficult. Dedistilled has made it on the very first try and makes it constantly. For me Dedistilled is godsent
@valentinkognito365 I did some quick tests with De-Distilled vs. Regular Flux.1 both GGUF and the Q8_O versions. ( https://civitai.com/posts/7985896 ). I'm truely amazed at how De-Distilled can butcher a simple prompt as "Flower in a green field".
I first thought i was doing something wrong so I went and downloaded someones workflow and I found out that this model is in some way very strange. The workflows I found are all bacis Flux.1 workflows with added overrides to simulate regular CFG. So this isn't a regulized model that can run on a Basic/Advanced KSampler what "CFG unlocked" will have you believe.
I ran all those tests when I was making my original workflow for Flux.1 Dev and no matter what you choose to impose over Flux to allow it to run Negative prompts only the "PerpNegGuider" node offers really good quality. So I don't know what the override nodes exaclty do that the De-Distilled Workflows all use but it's obviously not as good as "PerNegGuider". That combined with a Generation time of at least 145s. (on an RTX4080) while I can do regular Flux.1 Dev on my workflow in 45s. or about 100s. if I turn on Negative Prompts seems to be a very simple reason to let the De-Distilled model go for now.
I also did further research into the claims that it was trained on Flux Pro, but Pro atm only works via a paid for API Key so I don't know who originally trained this model as this is a forked model. But I can't find a stable sort of proof that this claim is just.
@Graybles there must be something weird. Can you give me the exact prompt and seed of that flower thing, I'd like to reproduce that on Forge
@Graybles I totally understand your frustration with the new De-Distilled Flux model not delivering the results you expected right away. This model is designed to offer enhanced capabilities, I can personally assure you of that please look at the images posted by @azimuthal and make sure Cross Post is enabled when viewing the images under the Q_8 model. It does require using a specific ComfyUI workflow to fully leverage its potential. From my experience, sticking to the recommended workflow—like the one I shared [here](https://gist.github.com/dasilva333/87bdd5b5b8ebba5515a9919ede0e3c05)—which includes the GGUF Loader, Dual Text Prompt Encoder, and KSampler with appropriate CFG ranges, makes a significant difference in the quality of the generated images.
It seems that some of the additional or unconventional nodes you’re using is most likely what's causing the subpar results for you, as this model is optimized to work best with the suggested setup. By following the tailored workflow and adjusting the CFG settings (3 to 10), you should see much better adherence and image quality. If you need any help setting it up or tweaking the parameters, feel free to reach out—I’m happy to assist and ensure you get the most out of the De-Distilled Flux model!
hi thanks for this model but i have a question what is the difference between fp8 and q4 models ? i have a 3080TI and 32Go ram, thanks :)
Q4 model will be smaller. The important is how much VRAM you have on your GPU. I've only tested myself the Q8 GUGG model and I have 16 GB VRAM (I could fit fp8 as well). Usually the difference between fp16 / fp8 / Q8 / Q4 / etc is neglectible. GUGG takes even less RAM as it is compressed. Under Q4 though, like Q3 Q2 etc it starts really degrading. Just try different ones.
@valentinkognito365 thanks a lot ! wich models do you think is better for my configuration ?
I don't know, try the smallest Q4 first then you'll see
The prompt adhesion is really VERY good ! Totaly worth the wait of the 60 steps. Does anyone know if the "Hyper" lora dev team are working on speeding up the process ?
I fully agree, and this along with the vastly superior quality, ruined distilled for me, it's over. I'm using flux since the very start. I could see the difference immediately.
These models are very recent I think, so we must spread the word around so that people start dedistilling even further and retrain the hyperdev stuff...
Was really excited after reading the description of your model. Unfortunately, using Forge with 8GB GTX1070, with GPU weight set to 6656, with t5xxl_fp8_e4m3fn + ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF text encoders and ae VAE, trying to generate 1280w by 768h image in 60 steps with guidance set to 12, with negative prompt, I got
"RuntimeError: mat1 and mat2 shapes cannot be multiplied"
The only change in settings from what I've been successfully generating with for weeks is guidance changed from 1 to 12, and steps changed from 30 to 60, negative prompt and of course the model.
Unsure what this error means and any insight you could provide would be greatly appreciated.
EDIT: Tried again with same settings minus negative prompt and got same error.
There are more and Smaller GGUF files available for T5xxl: https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main
Maybe with a smaller one your GPU would be able to function?
I'm using the same t5xxl version, clip and vae so I doubt this come from this. This could be you're using incompatible Loras. I had this error before when using the distilled models, so it's likely unrelated to dedistillation.
Have you tried setting "Diffusion in Low Bits" to "Automatic (fp16 Loras)" ?
Btw 1280 x 768 is going to be very long with it. I'm staying at 512 x 768 for now and it takes 1min30 average on my RTX 4090 with 16 GB VRAM. I found out the quality is surprisingly better than higher resolutions with distilled fp8.
(Please note I have only tried Q8_0 GGUG myself)
I am getting the same error. Tried suggestions in comments, but not fixed
The guy that made most of the quants versions is redoing them with a fix for Forge, I hope it should solve this. I just tried Q6_0 and I had this error as well. Stay tuned.
Can you provide the comfyUI training process for this model? It seems that the original process cannot run
Sorry I don't know, I don't use Comfy :-/
@valentinkognito365 How about for Civitai?
What do you mean ?
@valentinkognito365 LoRa training on Civitai. Not sure if one can select this as the base model to train on though.
@EricRollei21 No indeed you cannot.. I think you can only use the vanilla flux checkpoint online. Too bad, as for me these models are the most important thing to happen in the flux community since we're able to train and use LoRas
@valentinkognito365 It's on my list to get flux gym working
@EricRollei21 I have tried using them yesterday with kohya sd script (the bundle inside flugym) simply by replacing the flux dev .sft by it but it crashed - same thing for fluxdev2pro - at this stage it would probably be a good idea to contact the dev on github ; i know he's been working on adapting fluxgym for fluxdev2pro ; i've had the same error with both
this feels like a breakthrough and the "old" dev LoRAs seem to work with
For me it is absolutely a breakthrough, and the classic distilled Loras work even better
Thanks for making this. I am getting good results so far. Unfortunately I still get the famous flux grid of death (the vertical lines in larger images). Have you found any settings to lessen or avoid those? I have some workflow trickery that I can use to avoid those but it takes longer.
I didn't make it I just uploaded it here as soon as I tested it :-) Never had the grid. Use 60+ steps and Euler / Simple
I was getting this in Forge. Not sure if this will help but using UI mode (all) I made sure to slide distilled cfg to 0 on the hires fix area as well as the normal settings area and my results are artifact free now. 40 steps, cfg 3.5, hires cfg 3.5. hires denoising strength 0.35, distilled cfg 0, hires distilled cfg 0, hires steps 0, Euler/Simple, 1.75 upscale from 768 /1024, upscaler - ultramix balanced, (I kept all the CFG (hires too!) at 3.5, and the denoising strength 0.35 as well in order to establish a baseline but plan on experimenting.) These settings gave clean results with upscaling and no grid or other artifacts. I hope this helps someone. Please do experiment and report back any new insights.
I'm getting bad results when using this de-distilled model in an img2img workflow with the Shakker-Labs Union Pro tile controlnet. The image is smeared and with a cloudy texture even at 60 steps. Anyone else?
also bad results with standard clip,ae and t5xxl and 70steps
yeah apparently you need to use real CFG higher than 1
means you can't use your normal flux workflows
@sevenof9247 fyi read the instructions
@EricRollei21 I did use CFG higher than 1. Are you getting good results with img2img?
CFG = 1 means terrible results
What ui are you using ? Workflows can be tricky as everything so far was tailored for distilled models. When it works you will know it almost instantly. Distilled is almost helpless compared to what can be done on a complex prompt with loras using Dedistilled
@valentinkognito365 I use the latest ComfyUI version with all nodes updated as well. Maybe you are right. My workflow might need some adjustments. I will keep trying and will post here if I figure it out. Thanks all for the helpful suggestions. :)
would be nice to have a BNBNF4 version 11GB (clip, vae, t5xxl) all included
Can't run in Forge. All I get is
RuntimeError: mat1 and mat2 shapes cannot be multiplied (8064x64 and 256x768)'
I can use it with ComfyUI alright where it works quite nicely
I am having the same issue if anyone has a fix?
@Organoid There seem to be two ways to do Q8 quantization and Forge only supports one of them, I think?
I'd say to use the FP8 model instead, but for some reason Forge thinks it's a Schnell model and the output is broken.
@steamrick I'm using GUGG Q8_0 and the Forge log says it is loading a Schnell model but it works
The guy that has done the GUGG versions is in the process of redoing them with better compatibility for Forge. I will post them when he's done. Stay tuned
@jbar272101 Try the Q3 quant I just uploaded
I am trying to get this to work with Forge, and it fails with "You do not have CLIP state dict!" What is this and how do I fix it?
@CupcakeDreamer I'm currently trying as well, but I'm past that stage. Some models have Clip and T5 integrated, but for others you have to download them and turn them on in the VAE/Text Encoder section at the top. I don't remember the links for where they are though I downloaded them a while ago, sorry.
@valentinkognito365 getting error AttributeError: 'NoneType' object has no attribute 'sd_checkpoint_info' after generating one image when i try to generate more, the image looks fine in the preview, but always ends up as a solid black box at the end.
Edit: Nope, got it working! It was just my VAE was wrong.
@CupcakeDreamer this error has nothing to do with the model. Make sure you have the regular ae.sft (VAE), t5xxl file, and clip_l (or modified clip) loaded in the field at the right of the checkpoint
@Nukemouse the models on this page do not have anything else included. Also, make sure you select the flux workflow in Forge at the top left corner, you seem to be using SD
Please also make sure you are using "Diffusion in low bits" = Automatic (FP16 Lora)" in the top of the screen also, otherwise LoRas will probably not work or not work as intended. This apply to everything flux-related in Forge, not DeDistilled in particular
Using the recommended settings, fp8 and q3 produce blurry images when cfg > 1, regardless of steps or schedule type. This is the same effect I've had with other flux models, despite them being loaded as schnell. Q4 produces the mat1/2 error. Anyone else familiar with this issue?
Prompt adherence for the FP8 version I tried is no better than most FP8 models regardless of step or CFG numbers, in fact I found it worse than some. In terms of graphic quality, particularly in OOF areas its poor. It still suffers from the same concept and placement limitations as the original Flux model.
I feel bad that it's not working for some of you guys. I get totally opposed results. It just makes stuff that was nearly impossible for me with distilled, and most of the time on the very first attempt. Especially with LoRas, there is absolutely no comparison.
Examples :
- I have made a cum facial LoRa. Distilled : makes very little cum. VS Dedistilled showers the girl with cum and is able to make it appear just anywhere
- I have made a pov blowjob LoRa. Distilled : almost impossible to get the penis in the mouth despite extensive, desperate prompting. VS Dedistilled : penis fully in the mouth at the first attempt
I don't know about details, maybe I loose some, maybe be not. I don't care to be honest : It makes what I want it to make. I absolutely love it and now swear by it.
But then it's an experimental, proof of concept version and there's strong hope now that better dedistilled checkpoints will come out in the future.
I ve updated with info about Comfy workflow if you were using it. You might just not have seen the real thing yet. It would be impossible for you to say prompt adherence is not better otherwise. Its crazy good.
I've put homemade examples in the model description so you can realize what you are missing
I haven't tried the other quants, but q3 is super blurry on forge with the recommended settings. I kept the default forge resolution for FLUX, but otherwise tried adjusting CFG and steps to see if it could help -- nothing so far. Prompt adherence does seem pretty great, otherwise. I'll keep watching for updates.
@solss I'm not sure how to do the quants myself otherwise I'd happily do it. I don't think it's very difficult. So I'm waiting as well because I'd like to use Q4. With LoRas i'm a bit short on VRAM. Regarding Q3 I see what you mean. I got good results with it but then I got bad stuff recently as well (quality wise)
@solss Exactly, I get these blurry images on forge with any flux model which has cfg > 1 (at least fp8 or any quants, full always crashes). Perhaps a matter of forge version?
@firemanbrakeneck Check you have at least 40 steps and also you are using Euler / Simple, and Automatic (FP16 Lora) if using lora
@valentinkognito365 I have tested on varying amounts of steps (up to 100), low bits settings (yes, auto fp16 by default), samplers, schedulers (simple or beta mostly), dimensions, simple or complex prompts, with or without loras, you name it. Whatever was brought up in various issues on the repo in relation to this problem, but nothing makes a difference for me once the cfg is above 1 in flux (and for dedistilled, not even cfg 1 helps, I guess because it's ignoring the distilled cfg).
Also, I've tested the pipelines directly in the past through black forest and disty's code samples on some fp8s; even with a few steps (~10), you can already see an image forming, but with this blur, it seems stuck below even that threshold.
Looks really bad for me with InvokeAI even with the settings you recommended 😢
This is AWESOME. Total gamechanger. Thank you so much. Where can I learn more about the process behind this effort ?
I don't really know myself but you might find some here
https://huggingface.co/nyanko7/flux-dev-de-distill (author of the model) who apparently used an algorithm from a research article ?
And here : https://medium.com/@zhiwangshi28/why-flux-lora-so-hard-to-train-and-how-to-overcome-it-a0c70bc59eaf
That all I know. It's probably very technical...
Are you using Forge or Comfy btw ? I see some people very convinced like me and some have rather bad results and are skeptical (can't blame them if they didn't get results)
@valentinkognito365 I'm using Forge. The prompt adherence is so good. btw i'm using the FP8 version.
@chuckest It's crazy isn't ? I can't go back to Distilled except for drafting
COMFY USERS : I just updated the description with recommendations from @DaSilva who has been using the model for a while in Comfy with good results