PixelWave FLUX.1-schnell 04 - Apache 2.0!

@humblemikey I always glad to let everybody know about someone's fail...
[ JOKING ] [ SARCASM ]

@humblemikey THX

@humblemikey It does, and on a 4090 its around 6 it for 1024 px.

@vigilence Yea.. that's around 3-4000$ i do not have.. I'll have to suffice with my 4060ti with 16gb vram. Might be able to train a lora on this, if I can get these tokenizer issues sorted out in onetrainer.. I tried kohya a few months back and it just smacked me into a brick wall with OOM crap, sides, I like onetrainer's simplicity. Kohya is downright intimidating if I'm being honest.

Yea, tryin to figure out onetrainer myself. I got past the authentication errors by copying the folder from the official flux model, and just tossing the nf4 model in there (in place of the full-size model, which would take millionaire hardware to train), although now I'm at an impasse with the tokenizer. It expects a tokenizer for this thing, which I do not have, since that didn't come with it.. ChatGPT tells me I could train a tokenizer for it, but that just sounds like a bit much headache..

@humblemikey Ok, thanks. I tried I still have artifacts, well ok, for lore will be other models.

@humblemikey I tried this but it doesn't do much on my setup. I guess I'm doing something wrong

@humblemikey compare same prompt with stock dev and this model , there is detail drop,stock flux is pretty crispy but this one kinda jpegy

@niccc could you provide examples?

talk about entitlement , too much for you don't use it . truly entitled , the dude spent 5 weeks on his local gpu to give something , and you come here like you paid for his bills ,rig and effort . my god ...

I've also noticed rather serious quality degradation if using default Flux-dev settings (Euler, 20 steps, Normal scheduler) and result does look like an overcompressed JPEG. interestingly enough generation is also faster than vanilla Flux-dev.

however, switching to SGM Uniform scheduler brings quality back to normal on all samplers I've tried.

not sure about op, i'm using latest SwarmUI and Q8 model variant.

@caki dood just cause i pointed out that photo style is not as sharp as stock does not mean the model is trash, are you crazy ? Is it just black or white with you.

@randomkcoala3ill try it, i used more steps and quality improve but still its not as crisp as stock flux 

@humblemikey Did you download it from civitai? Because we're all downloading it and trying to unzip it and getting errors when unzipping it.

@psspsspsspssspss you do realise you responded to the author lol... geez this site and its users I swear.

@humblemikey Thank you, now it works fine. : )

@axicec The irony is strong in this one...

@humblemikey thanks!

Same here, I can't get it to work with LoRAs for whatever reason, so it's useless to me.

Not sure if they had been there before, but the author added GGUF links in the description

same problem with Loras, otherwise it's amazing when using it stand alone

Same here, I can't use it with LoRAs either, I thought it was just me, but it's basically useless without LoRAs.

I would love to know the training config/method as well. I have been trying on and off with my 4090 but struggling to get good results. I saw limiting it to one concept/batch of images at a time so was going to look into that aspect. Would love more information though on the settings being used (and which platform).

I would also like to know how to load it in SwarmUI, I would like to try this model out.

Which exact model did you download? I'm about to try the Fp8 safetensors version in Forge, so it would be good to know if that's the issue, or if you're using one of the GGUF's. I'll let you know how I get on.

@wideload I am using the FP8 safetensor in Forge and it is working superbly. With HiRes Fix I have found Latent (antialiasing) as the upscaler is giving the most pleasing results.

@charidot113 I agree. Just tried the same and the results were spot on. Tested photo, watercolour and it's just finishing a pencil sketch. All three exactly as I would expect.

@wideload 
I used the one which is 11GB large ... gives me a lot of incoherent results - sometimes it works better if I increase the CFG (which you are not supposed to with FLUX, are you?), sometimes it starts to burn it with CFG > 3 already ... And why it struggles with text and hands ... ?? I am quite curious what your experiences will be. Hope you're going to share ... At the moment I am still reluctant to trade the benefits from this FLUX art model which on my 4090 still takes 45 seconds for an image against SDXL which is more than 4 times faster ...

to me specifically adding 'photo' or a 'movie still' etc. helped

@sunflower96733 I'm running it in Forge and it's producing fantastic results. I put it in the standard checkpoint folder, have you tried moving it from the UNET folder? I'm also using the FP8 safetensors model from the description which is 11.1Gb. It's worked with Euler Simple, DPM++ 2M SGM Uniform and the Forge Realistic samplers so far. CFG was 1 and Distilled CFG was 3.5 with 20-30 steps. Resolution was often 896x1152 but also higher. Hands have been perfect every time for me and the only issue was using loras of people. For them I had to drop the lora value down to 0.9 or it appeared blurry and pixelated. Multiple styles have been tried, including the three mentioned above and they all worked very well.

I don't use or really know much about Stability Matrix so I don't know if that's an issue. If you have access to Forge or ComfyUI (outside of SM) it might be worth trying them as the model obviously isn't at fault.

@wideload 

I already got the confirmation that other people also have problems when using it in Forge. Even if I copy the same parameters set from Civitai the results something are very similar, and sometimes not.

What VAE/Text Encoders did you use? And is their sequence of any influence?

@sunflower96733 I use ae.safetensors, t5xxl_fp16.safetensors and ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors. Those are my standards and I just tried them in two different sequences for you and it changed nothing.

The only other setting I haven't mentioned is Diffusion In Low Bits is set to Automatic (fp16 lora). Beyond that, I'm drawing a blank on what it could be.

@wideload 
Are those encoders here really make a difference?
https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main
And which one did you chose and why?
I am currently using the t5xxl_fp16.safetensors, or the fp8 version - but with my motives this doesn't seem to make a big difference ...
By the way: does emphasizing with this syntax (hat:1.5) also work in FLUX? It didn't seem to do much when I tested it with this model here ...
Did you also check text generation? If I leave CFG to 1 text is practically unusable. Only begins to make sense with CFG 3 or above, also massively depending. But even if the text is almost right, it still struggles with names. Can't spell 'Cathy' - makes at least 'Cathah' out of it.

@sunflower96733 I don't know how much of a difference it makes. I chose it because a video I saw recommended it (by Olivio Sarikas I think) and it hasn't done anything bad so I stuck with it.

The one I have appears to be about half way down that list, it has the exact same name as I posted and it's 323mb. It goes in Forge's model/text_encoder folder.

I also tried the fp8 version of t5xxl and it made no difference.

Changing the value with that syntax sometimes works and sometimes not IMO. Or it's effect is weaker than the SD models.

@sunflower96733 I only did a couple of tests with text and just like the base Flux it isn't perfect every time but it still works well most of the time. I usually find if the first generation doesn't work, the second does but I'll do some more testing on it when I get chance.

Are you still only getting 'impressionist' style images or have things improved since your original post?

@wideload 
The issue with the model not respecting well known art styles as good as expected I could improve with setting a higher CFG value. But this can only go so far.

@sunflower96733 I don't know then. I'm not negating your experiences but I haven't experienced a single one of the issues you have and I haven't changed the CFG once. It's working fantastically for me, with the exception of some character loras I tried.

Maybe it's the particular art styles you are trying to recreate, it could be anything at this point if your settings are all correct.

@wideload 

Yes could depend on my art style and content of the prompt. The interesting thing is, that this model here doesn't have the exact same biases against NSFW e.g. as the original has. I will make an in-depth test with some styles and artist names - but not being able or recommended to up the CFG scale is of course a severe limitation of FLUX, since there is no other parameter to compensate for this. If I want my image to be 'cubist' of course there are various levels of cubism. Many users seem to stick too much with 'realism' and don't venture deep enough into abstract fields to even notice this...

And what about the (shoes:1.5) way of upping things? Did you test this?

@sunflower96733 You should be able to change the Distilled CFG in Forge (default is 3.5). I believe that's what you change for Flux but I've never needed to do it. Don't take this as fact though.

Yes, NSFW is much better with this model. It also does fangs out of the box without lora. I tried a vampire seeing as it's spooky month. Haven't tried blood and gore yet.

I had mentioned changing values but I guess it was easily overlooked as it was at the bottom of the comment. Changing the syntax values sometimes appears to work and sometimes doesn't. It may be that it's effect is weaker than with the SD models. I'm not too sure to be honest but I tend to prompt for exactly what I want with Flux as it seems to work better.

I'm one of those realism people. Sometimes I'll do sketches, watercolours and a few others but never something like cubism. It's not my taste. Of course you could train a lora yourself and merge it with Pixelwave, I think that's possible.

@wideload 
I increasingly get the impression, that for my take on AI art, FLUX has more negatives that outweigh the positives. Realism is totally not my thing - to my opinion, we should leave this to the photographers as long as possible. And everything FLUX is said to do better than SDXL is not reliable (prompt adherence, text rendering, good hands), can be easily fixed in SDXL with outpainting or regional prompting - or has massive downsides that come along with it (only half baked NSFW, if any). And the rendering takes MUCH to long, even on my gaming laptop 4090 with 16GB VRAM.