100% better. im blown away. the loras i trained for flux 1 dev are almost obsolete. its incredible how good the images turn out without any lora.

With the gguf Q8 it's definitely better, less fluxface for whatever its worth, haven't hit this fp8 yet

Flux Dev is only as good as the user, as tends to be the case with all AI. AI magnifies skills, and it magnifies lack of skills. The pity is that the resources exist for free for anyone to up their game.

It's definitely more realistic than any other open source model, when it comes to be similar to real photography. For people who knows photography and overal lighting and film it's the best model because it really has a good understanding. But if you're business is creating waifus, nsfw and anime, maybe it's not your best option.

way less body horrors, and faster at least on FP8.

denrakeiw 
20B is way too much.
Why nobody is making 6-8B DiT unrestricted model with proper training?!

fp8 requires around 12gb VRAM

JunkieMonkey69 No, no it does not. Only LLM models should reside in VRAM if possible. For video ALWAYS and for image for those without an excess of VRAM, a model should always stream across the PCIe bus as needed. In theory a model could use ZERO VRAM, and still work full speed, but ComfyUI is too badly coded for this perfect situation, so some VRAM is needed for streaming- but a fraction of the actual model size!

blobby99 so i can run the full fp32 models on my rxt 4070 super?!?

JunkieMonkey69 I just ran the original release KREA file from BFL (22GB) with my 12GB 4070TI. but its much slower as if it would fit completly.

no.

This quantization won't fit entirely on 6gb of VRAM. However there are a few quantizations on huggingface that would. There are Q2_k, q3_kS, q3_KM https://huggingface.co/QuantStack/FLUX.1-Krea-dev-GGUF/

I'm seeing 16-19 vram usage. I hit 24 vram using 2048 size.

JunkieMonkey69 Here I got error only when trying to load more than 32GB data (Model+Encoder+Clip) with only 8GB VRAM, but 32GB RAM.

You can pass to your download manager a link with the model id number (which you can read from the standard download link, for this model it is 2068069) and your api token, something like this :

I always use this to download large checkpoints.

no.

yes.

yes

yes

It works with Forge. I haven't found the best settings yet, but with initial testing, I've found the model to be very accurate in following the details entered in the prompt. The ages of the subjects are perfectly reproduced, in my case I had to reduce the weight of the details on the skin because I see that the model is very accurate and can exaggerate with the details. The Lora settings could be excessive in many cases.

have you tried the (euler ancestral/ipndm + Beta) combo?

JunkieMonkey69 aha! Now that's blur! Thanks for bringing up this combo, res_2 gives a much sharper output.

JunkieMonkey69 by the way, your loras are amazing. Great work!

Primaveri thank youu

LoRAs trained on flux or dev2pro may not work drop-in style for Krea. At a minimum, it seems like the strength needs to be dialed back some from what you would normally use.

Dark_infinity Thanks for the reply, so how do you recommend I proceed?

darkness1it try them with a lower strength (if you use 1.0, try 0.8 and see if it improves). I'm still investigating other ways of adapting them.

fp8 3-5% better, very minimal, but uses less vram

if you can run fp8,never run Q series

Minagi any reason? cuz I have download Q8.... very sad.. I have to re-download fp8

zczcg q8 requires a bit more vram and the quantization process for the q8-Q2 models are less accurate than just the simple fp8/fp16 models

JunkieMonkey69 Got it. I have a small question: And why they offer the Q8 both fp8 version instead of only fp8

zczcg different people work on different models. gguf models are very popular with llms. and with gguf models you have the option to make q2,q3,q4,q5,q6,q8 of the same model, which enables even lower vram tier users to use the same model, while fp8 just halves the vram needs

JunkieMonkey69 So they just use it to make q2.q3.q4....cuz i only care about the generate time and the output quality.

I know you're asking about the quantized version but I'm also going to chime in here that I like fp8 just as much or even more than fp16. It generates faster and saves on disk space as a bonus too. I typically always use fp16 but this fp8 is growing on me.

HackAfterDark it`s true, fp8 faster than Q8 5 times. I don't know why

this Q gguf madness has become a religion. Almost everything else is better and more performant than gguf: int4/fp4/fp8/fp16.. People think that gguf is a hammer and everything is a nail.. People need to start thinking independently and not to mindlessly go with the masses

zczcg It's LoRAs. The models alone should be similar speed, but iirc, GGUF internally is blocks which while it uses more VRAM, offloads easier. These blocks are slower to apply loras to since it will span boundaries or something.

they kind of half-work. it depends on the lora--you just need to test to see. increasing the strength doesn't help them come through more it just leads to glitches. loras will likely need to be re-trained.

Some of the LoRAs I've trained do work with Krea and make it even better. I'm in the middle of a bunch of comparisons right now. I will try more LoRAs over time, but yes, generally they'll work.

It seems that it's just Flux Dev with additional tune towards realism and removing the Fluxface like any other photorealistic Flux-based checkpoint presented there.

mphobbit it feels a lot more like flux pro, what flux always shouldve been, and also feels a lot like some of my photography loras--like, just training a photography lora on original vanilla i feel like unlocks parts of the model that are also accessible with flux krea. like its a lot more responsive to prompts. i have no idea what they are doing under the hood but it feels like krea is just BFL giving the community more of what flux pro is, but not completely

cutetodeath78409597 I haven't used Pro much but I'm finding it's not bad for skin tones and softening the face. It does remove some of the plastic/airbrush/gloss look and helps with lighting, but I'm not finding it picking up on details in hair, eyebrows, etc. Not sure if it's adding or taking away from anything either to be honest. I'm still using my photography LoRAs which improve it further. So far it's not doing much for me, but at least it's not worse.

some of my loras work better on this model while others are absolute poo poo. they also trained the models prompt interpretation .

JunkieMonkey69 Yea, I think mileage always varies with LoRAs in general though. So I'm not real surprised. You're right about the prompting though, I definitely think something is different there. Not sure if that's good or bad, maybe just different and something new to learn.

Be that as it may, it is possible to train with this model, but the only LORA I have done so far has flaws, especially in the hands, which come out broken most of the time.

JabGal Yea, that's a good point - I've mostly been generating portraits that don't include hands or other details so I'll have to see. Also, I do want to train some models on it and see what happens. I'm still heavily experimenting and testing, my feelings on the model may change of course.

HackAfterDark I actually did a 25 seed test with original flux+one of my photography loras and the responsiveness to the prompt is about the same if not better than krea, and the original flux+lora pictures look very similar to both vanilla krea and krea+the same lora... so i dont think they trained the encoder or anything to do with prompting, its just that original flux is 'dumbed down' in some way perhaps with distillation. when comparing all of them i prefer original flux+lora the most, at least with this particular prompt. maybe i'll do an article here to show my work, if the site ever returns to full functionality lmao

cutetodeath78409597 cool, good to know. Maybe I'm imaging things here, but I also think the fp8 version of Krea is better than the fp8 version of Dev. I'm finding it a little easier to work with and so while I'm not really finding it better OR worse than Dev here...if I can gain some speed and save on disk space? That's a good win.

HackAfterDark HackAfterDark for sure. I'm definitely going to keep both around, I feel like both will be good for certain things. I do think Krea follows the trend of 'better=seeds look similar' that has been a symptom of both hidream and chatgpt image gen. original flux + lora has the most wild seed variation, which is something i value a lot, i dont want to see the same pic over and over again, thats what i love about SDXL is it farts out a lot of trash but then every once in a while it gives you a seed that makes you rethink the prompt, its more of like an artistic partner in a way if that makes sense, whereas 'better' models just return something thats way too literal and boring. its like when you train a lora too long and it puts everything to the 'mean'.

For me it Hires Fix/Refining that kills it. For whatever reason it subtracts detail instead of adding it. I opened a ticket on their HuggingFace: https://huggingface.co/black-forest-labs/FLUX.1-Krea-dev/discussions/5

denrakeiw thanks, I'll give it a try

denrakeiw I appreciate it. I'll check it out.

Normal ones works great.

ty! I was thinking if it was breaking some policies that I don't know yet. 🤦‍♂️