The FP8 version has been uploaded!

Try lowering the weight of your LoRAs. I’ve found that this model understands LoRAs better than any other model I’ve used, it’s epic. However, sometimes it understands them too well, making any defects or signs of overtraining very obvious.

What works best for me is either decreasing the weight, increasing the CFG value (if the LoRAs seem to have no effect), or lowering the CFG if the effect is too strong. If you have an earlier training epoch, that might work better too.

Some of my LoRAs require half the weight I was using with regular Flux-dev, and for some style LoRAs, I even need to lower it as much as possible (sometimes down from a weight of 1 to 0.1) to see the intended effect.

You’ll need to relearn how to work with LoRAs for this model, but it’s definitely worth it. Also, make sure you’re not using Flux guidance.

@sparkleboy599 Well yes and no, the real answer is that flux-dev is not so good at training and capturing details. It's better than most, but it does mistake. Sadly we're only finding that now that dedistilled models are coming out. Dev2Pro is one of those and it will do an amazing job compared to dev. But, does it mean we'll have to redo everything? Probably not, some stuff will be salvagable... some is just plain better, and some will have to be retrained.

@sparkleboy599 Personally I would use the Dev2pro. My model is an hyper version based on the same idea. Mine limits the number of steps needed for the model to converge; but, for training I prefer more steps with a lower learning rate.

what are you using, comfy, forge or something else? From what I’ve seen so far, sometimes it doesn’t do anything and sometimes it affects the image in a bad way. I believe that it is best to disable it, but if you do use it, it won’t change much.

@wregiszter632 I haven't used forge in a while, but I think you can load it as flux schnell and still use Loras. I believed doing that it will simply remove the flux guidance slider. If you try it can you please confirm if it worked for you and if loras are still working ?

@wregiszter632 Then that’s the way to use it in forge, thanks for the reply! 🙂

Swarm auto autodetects and downloads/load the clip and vae models. That’s why.

This is a pruned model, so no it’s not included. I will add a link for the T5, my favourite clip_l, and the vae from Black Forest Lab in the description. Now for quality, disable the flux guidance, and use a cfg value of about 1.5 for the 8 steps model and a cfg value of 2.5 for the 16 steps model. Those are only recommendations you can always adjust those values to your liking.

As with Stable Diffusion models, lower values give more realistic results and higher values outputs more artistic results.

Use fp8 and load it as schnell. If you do so, the only difference is that it won’t load the flux guidance function, which you don’t need. This model is based on Flux-Dev but its limitations have been removed. Which means it takes more time to generate better images if you put time into it. The reason behind that is that it needs more computing power with a CFG value greater than one (the guidance was implemented to counter this).

Now for the text encoders and VAE, I will surely add this to the description. Thanks for the idea! Not everyone knows what a pruned model is and I should have think of that.

Do you believe it would help other Forge users to have a full model with everything baked in? Personally I find that limiting in comfy, but if it helps you I will upload one. :)

@Satsuma9977 I will add the fp8 models with everything baked in during the weekend :) Have you tried to make it work as schnell ? Another user reported that it worked well for him this way in another comment. One more tip, you should lower the weight of your loras if they don’t work well anymore. The model understands them better and signs of overtraining might show up at their normal value. Lower that and you should get awesome results!

Yes you can use LoRa as with any flux-dev models!

@amandalevres I totally agree! Usually, I use the regular Flux-Dev for rough drafts, and once I get the prompt right along with the settings, I switch to Hyper-16 dedistilled for the final draft. Interestingly, with Hyper-16, you can use the same parameters for everything except for the guidance. This last part is somewhat equivalent, though, so distilled guidance ≈ CFG value!

@Alexatr use CFG instead of distilled CFG if you can. Otherwise you have two options, load the model as schnell or leave the distilled cfg to any low value. It doesn't do much anyway.

Depending on what you use : chenge the nodes you use for comfyui, disable distilled cfg in swarmui, or load as schnell in forge

I've manage to get great results with an RTX 2060s with 8gb of ram. I've found a way to generate at 2k with 1 to 2s/it There is some quality loss but most of the time it's barely noticeable. PM me if you're interested!

I'll try to upload it as soon as I can! but for whatever reason uploading on civit fails most of the time.. It will be there eventually.

I've stopped usinf gguf models lately (updated my video card). However if you havea ggug loader installed, just drag and drop the following image in Comfy https://civitai.com/images/65675938. Replace the "load diffusion model" node with the gguf model loader node, then set the proper parameters (using the recommendations depending if you use 8 or 16 steps model).

Personally I use SwarmUI, it's simpler to set up, it installs all the requirements for you and you still have acces to comfyui for more diffucult task.

Try to load it as schnell in forge.

A lot of people had issues with that ui and dedistilled models. I don’t know why, but it seems to be forcing distilled cfg on anything loaded as dev (comfy and swarm ui handles this way better) but that simple workaround should do.

As for the number of steps, it’s been a really long while since I’ve used this model but try using 24 steps (using more steps will not take necessarily make it longer) and a cfg of 2.5 with dpm++ 2m and beta. That set up without Lora’s should do wonders and you can always adjust the setting slightly once you have something that looks good (less cfg=lower prompt understanding but more realism).

Then if you want to use Loras, add one at the time. The main issue with and distilled models is that it understands Lora’s way too well. That means that any defect, overtraining, or weird stuff included in a Lora, which would work normally with dev with some occasional hiccup, will be horrible when paired with a dedistilled model. It will pick up any defect and expose them.

The solution is super simple add one Lora at a time, use the recommended strength, and if you run into issues lower the strength by increments of 0.1. Generate your image and then repeat if necessary. Distilled is smarter than dev it doesn’t need the same strength than dev.

@jj05690569159 perhaps it already is, but I believe it’s close to the drop down menu where you load your model along with t5 and clip_l at the top of the ui. Or edit the metadata of the model within the ui, if that’s possible.

sorry if I can’t help you more than that with this part… as I previously mentioned I’m a swarm/comfyui kinda girl ahah

The best advice I could give you is to try out swarm, it’s super simple and a lot of people who tried distilled migrated this ui since it’s way better at handling complicated task like using dedistilled (and many more things, it’s like the beginning of what you can do with diffusion models and you will not be able to go much farther with forge at some point).

@jj05690569159 I only use my other model Pandora https://civitai.com/models/1143800/pandora-or-flux , I’ve made it for myself! And I’m sharing it here. It’s wildly different but much much better. There’s two version Dev and Pro, the pro version is a dedistilled model. However at this point, the latest dev (beta) version gives results as good as my latest Pro version (alpha)! I’ll update the later soon, but I’m still testing it.