does the p40 do better at bf16 than fp16? bf16 is a truncated fp32, right?

Sheepishly bringing my tech support issues into the new thread >>101996387

File: file.png (1.58 MB, 784x1264)

1.58 MB PNG

The tarot card LoRA is kino

File: 2024-08-20_00401_.png (1.12 MB, 1280x720)

1.12 MB PNG

>>101996391
thank you baker

>>101996415
Try the Load Diffusion Model node instead of Load Checkpoint

File: 1702801896081651.png (1.27 MB, 1024x1024)

1.27 MB PNG

somewhat lewd armor paladin miku, wow ingame screenshot (prompt)

>>101996415
Try loading a sample workflow

https://comfyanonymous.github.io/ComfyUI_examples/flux/

I have a flux lora training question. Am I supposed to use this T5 attention mask option in kohya? Seems like it's not the default. I see some commits in SimpleTuner that added something similar.

As I understand it, every T5 prompt is padded out to the full 512 sequence length via the tokenizer API. Attention mask just zeros out the vectors corresponding to the padding tokens (what even are those vectors to begin with?). I guess it matters how flux was originally trained and what happens at inference time. Does anyone know how comfy / forge handle the T5 prompt? Does it use the tokenizer padding, or zero out those vectors explicitly?

File: ComfyUI_06622_.png (928 KB, 1200x768)

928 KB PNG

Base model itself understands a bunch of video games.

>>101996539
should have censored that pic tbqhwyf

File: ComfyUI_04961_.png (1.25 MB, 1024x1024)

1.25 MB PNG

Finally, Miku y2k
https://civitai.com/models/667307

File: ComfyUI_32624_ copy.jpg (828 KB, 1024x2048)

828 KB JPG

>can we have this Miku pizza?
>we have Miku pizza at home
>Miku pizza at home:

Is this the blessed bred?

File: ComfyUI_00431_.png (1.9 MB, 1024x1024)

1.9 MB PNG

>>101996696
maybe

>>101996701
what's the most impressive about this model is the level of details in far away distance, it doesn't look like a cluster of mush anymore like we could see on the SD models

File: ComfyUI_32627_.png (1.18 MB, 1024x1024)

1.18 MB PNG

>>101996696
Now it is.

>>101996435
bruh this is a blue board

File: fs_0298.jpg (618 KB, 4096x2560)

618 KB JPG

>>101996621
>>101996640
dig it

>>101996986
S
E
K
K
O

File: ComfyUI_00939_.png (846 KB, 720x1280)

846 KB PNG

Letting joycaption handle the description will give you flawless results

File: ComfyUI_00062_.png (2.81 MB, 1920x1080)

2.81 MB PNG

>>101996948
kino

File: ComfyUI_00940_.png (809 KB, 720x1280)

809 KB PNG

>>101997091
All you have to do is change the small details or smash multiple descriptions together and you can get what you want easily

https://reddit.com/r/StableDiffusion/comments/1ex3ol5/comfyui_experimental_rtx_40_series_update/
For the 40xx card gigachads, you should see a significant improvement speed with this flag:
--fast

Still running a 8GiB 2080 (non-Ti) and trying to get flux working. I've downloaded stuff into the right directories and tried to get it running, but it failed with "mat1 and mat2 cannot be multiplied" point at it being the different models being used.

Has anyone gotten a 8GiB card running? If so, which files are you using?

File: file.png (2.34 MB, 1024x1024)

2.34 MB PNG

>what's the most impressive about this model is the level of details in far away distance, it doesn't look like a cluster of mush anymore like we could see on the SD models

>>101997278
I'm running flux1-schnell-Q8_0.gguf and t5xxl_fp16 on a 3GB 1060 using forge. The rest are black forest default files.
It's slow as hell but it runs. So you should be able to run it for sure.

It's so fucking slow when you use a lora... goddam I wish flux had more concepts in it, maybe someone will fix that with a giant finetune, my Hopium is ready

File: ComfyUI_temp_foads_00052_.png (1.34 MB, 1024x1024)

1.34 MB PNG

>>101997202
4060 Ti and I see no noticeable improvement with GGUF model

>>101997374
4060 ti and went from 2.4s to 1.8s per iteration

>>101997387
with the gguf loader?

>>101997395
forgot to say that no, no gguf

The BEST lora has come out, only for people with refined taste. A Wong Kar-wai cinematic lora!

https://civitai.com/models/667594/wong-kar-wei-cinematic-style-flux?modelVersionId=747253

File: file.png (2.49 MB, 1024x1024)

2.49 MB PNG

>>101997360
Are you filling up your VRAM and spilling into RAM? Try a smaller quant if the slowdown is significant, it might be the case.

>>101997432
I should see that movie, I know it from there kek
https://www.youtube.com/watch?v=r2env-txnms

n slur

>>101997374
>I see no noticeable improvement with GGUF
it currently only does faster fp8 e4m3

>>101997445
no I have a 24gb vram and the Q8_0 is only using 15gb of it, it's just that loras are slower on gguf because of some complex dequant math shit that's in there

>>101997432
https://www.youtube.com/watch?v=ZGZGNrGyo2A

>>101997479
Well, GUFF backwards is FFUG after all.

>>101997456
Nice

>>101997488
Lol it captures the style well

>>101997510
this man is spitting facts

>>101997510
Oh ffug.

>>101997466
f slur

>>101997432
>580MB
bloated piece of shit

File: 1711947212731174.png (169 KB, 1546x830)

169 KB PNG

>>101996476
>>101996555
Followed the flux setup, downloaded all the required CLIP and VAE stuff, same error. I am CURSED!

>>101997551
redownload the NF4 checkpoint

>>101997488
I want every single movie and TV show (including Japanese, Korean etc) all built into one finetune.

File: file.png (2.23 MB, 1024x1024)

2.23 MB PNG

>>101997551
>>101997568
>nf4
why tho

>>101997551
that's because you're loading a nf4 model with "Load Diffusion Model", that's not the good node, you should use this
https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

and don't use nf4 it's shit, use Q4_0 instead, better quality + same size
https://github.com/city96/ComfyUI-GGUF

File: file.png (2.46 MB, 1024x1024)

2.46 MB PNG

>>101997569
>I want every single movie and TV show all built into one finetune.

>>101997583
what's she gonna do

>>101997569
>I want every single movie and TV show (including Japanese, Korean etc) all built into one finetune.
the first mf who makes a finetune that has all the movies/games/anime character + celebrities will be a true hero, forget about NFSW, making Hatsune Miku playing ping pong against 2B with Alex Jones as a referee in the style of Greg Ruwotski is the true goal to achieve

>>101997545
I mean, people make 200mb sdxl loras and no one bats an eye, 500 is comparatively small (flux is 23gb, plus the 10gb of t5...)

>>101997641
>no one bats an eye
I do, those are bloat too

File: 1715278455669445.jpg (250 KB, 1280x720)

250 KB JPG

>>101997568
>>101997579
>why tho
I was told it could run with 8gb of vram. The full dev model makes my computer into the mustard gas

>>101997581
alright... back to downloading...

>>101997651
based

File: 2024-08-20_00564_.jpg (385 KB, 1440x2560)

385 KB JPG

>>101997202
can confirm it works, quality cost was visible tho, see pic related (without) and follow up (with --fast)

stats without --fast
>1280x720 22 itertations in 14s, 1.56it/s
>upscale tiles 768x768 between 2.05it/s and 2.56it/s

stats with --fast
>same res 22/22 [00:11<00:00, 1.88it/s]
>upscale same res, between 2.94it/s and 3.62it/s

follow up post for quality comparison

(all done on a 4090)

>>101997677
too far, anon

File: 2024-08-20_00566_.jpg (369 KB, 1440x2560)

369 KB JPG

>>101997689
pic related with --fast, atleast for this picture the quality loss it extreme to bother with --fast

also dont forget you can only do this on fp8_e4m3fn

File: file.png (2.18 MB, 1024x1024)

2.18 MB PNG

>>101997610

>>101997472
This fails for me, looks like a CUDA error, what version should I have?

File: file.png (2.13 MB, 1024x1024)

2.13 MB PNG

Worry makes the chin worse.

>>101997598
She wants to know how much your Eggplant costs so she can take it home.

>>101997610
Yes! you get it!

>prompt "This is an anatomically correct image of the inside of a vagina taken from the uterus with a specialized type of medical camera."
>it has teeth

>>101997773
Try something like small nose and big lips or something, it seems like it tries to make a different face

File: ComfyUI_temp_ostyt_00002_.png (1.1 MB, 1024x1024)

1.1 MB PNG

>>101997711
Quality first. This model can give you what you want with a few tries, speed isn't so vital when it nails it.

>>101997808
it really sucks at kissing and facial expressions more generally

File: file.png (2.65 MB, 1024x1024)

2.65 MB PNG

>input random words in another language
>get pretty gens

>>101997792
>He doesn't know

>>101997792
https://www.imdb.com/title/tt0780622/

File: file.png (2.43 MB, 1024x1024)

2.43 MB PNG

>>101997851
I know. The HR Giger pill is the hardest pill to swallow.

>>101996417
lol that looks adorable

For vramlets, is GGUF or NF4 better?

>>101997954
Q4_0 is better than NF4 (both have the same size)

File: file.png (2.7 MB, 1024x1024)

2.7 MB PNG

I feel like I'm spamming the thread, but these are too cool.
>>101997954
GGUF for everyone although NF4 can arguably be faster?

>>101997964
Seems like it goes slightly over 8gb when running, do I gotta go a little more quantized than 4_0 or is there some other setting I can fiddle with?

File: file.png (2.2 MB, 1024x1024)

2.2 MB PNG

There we go.

>>101997985
there's smaller models than Q4_0 but I'm not sure if the quality is better or worse than nf4

>>101998002
who am I kidding, I've already comitted to this download...

File: ComfyUI_temp_ostyt_00006_.png (1.25 MB, 1024x1024)

1.25 MB PNG

>>101997833
Flux can make two characters kiss reliably, to me that's a lot.
I haven't tested expressions too much, but I haven't noticed anything worrysome.
What are you comparing it too exactly? I mean it's not perfect but against reality every other model is going to be inferior for a while.

>>101998037
>Flux can make two characters kiss reliably, to me that's a lot.
It's the kind of kissing where the faces just touch at the lips, there is no head tilt, the lips aren't puckered, etc
>What are you comparing it too exactly?
DALL-E 3

File: file.png (1.84 MB, 1024x1024)

1.84 MB PNG

I think it's done.
>This is a painting by HR Giger. It features black and white airbrush forms that convey an oppressive and surrealist atmosphere, blending anatomical horror with Freudian ideas about sex.
>The style is realist. Dark. Black charcoal. Existentialist.
>The inhuman eyeless creature has bone and rock growing out of its orifices, teeth were none should grow, and inverted horns.

Being able to iterate fast with Flux is so nice.

12gb lora dude, are you using xformers?

File: ComfyUI_temp_ostyt_00012_.png (1.23 MB, 1024x1024)

1.23 MB PNG

>>101998062
DALL-E3 allows kisses! Scandalous! bring out the dog! What is going to be next: hand-holding!

>>101996435
Reported this for nudity

>>101998178
They aged them closer tho

>>101998206
I wish it understood cyrillic I would love to so some soviet propaganda posters

Why doesn't anybody here do video?
>>101998240
You can use Google translate and Photoshop for that.

>>101998178
>DALL-E3 allows kisses!
Absolutely not, not anymore.
Point is it is evidence these models can have great control over facial expressions, better than what Flux allows.
Just another thing to curse BFL for while impotently shaking our fists.

>>101998240
couldn;t one use a translator (online?)
I can imagine the english-russian translation is decent enough?

>>101998251
no good local video model

File: bComfyUI_104799_.jpg (1.15 MB, 1440x2048)

1.15 MB JPG

>>101998073
is that how you have to prompt now? i haven't messed with this stuff since mid last year but flux brought me back.

File: 00006-3567203612.jpg (336 KB, 1552x1200)

336 KB JPG

>>101998302
Wow

>>101998255
I find it fascinating that you've been with this for days.
>>101998286
Imma try svg

File: ComfyUI_32655_.png (1.09 MB, 640x1280)

1.09 MB PNG

>>101998125
Yup, I installed these versions of torch and xformers. I think they're enabled by default, so you don't need to use the --xformers argument.

pip install torch==2.4.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu124
pip install -U xformers --index-url https://download.pytorch.org/whl/cu121

File: 2024-08-20_00584_.jpg (406 KB, 1440x2560)

406 KB JPG

>>101997202
>>101997689
>>101997711
slight correction to this

while there is a small quality change with --fast the biggest impact was the change in comfy code that loads lora, the version I still one had to load loras had heigh weight, so that changed the output, cause the same prompt with same lora weight was over emphasizing the lora, when lowering the output with --fast is way more similar to the original picture

although the ufos are now lamps, kek

>>101998324
*svd
>>101998302
The text decoder for Flux is a LLM that works with natural language

>>101998329
something's fucking up and it won't tell me what, it just says it raised an error, no shit.
is it normal for xformers to downgrade your torch version? or does it not matter

File: 1697523294823924.jpg (167 KB, 800x1170)

167 KB JPG

>>101998240
Just prompt harder and add text using ps

File: photo00015.jpg (131 KB, 1464x1064)

131 KB JPG

>>101998324
I dont really care

>>101998362
I think having models do text, while impressive, it's a waste of resources at the level we're at.

>>101998360
I dunno.

>>101998324
>Imma try svg
SVD sucks, no conditioning besides a vague "amount of motion" parameter and CFG
there were a few proof of concept tools that allowed to select which parts of the image should move and it worked but made gens even slower and lowered the resolution
best it will give you is a slight movement of the camera or the subject, anything more than slight will cause defects to show
Gen-3 and Kling are really ahead in quality. Sora is technically the best but only a few people outside OpenAI got to use it for a few projects.

File: ComfyUI_04988_.png (1.25 MB, 1024x1024)

1.25 MB PNG

Flux barely knows Chun-Li, that sucks

>>101998393
It's probably helpful to have a layout designed to have X amount of text, even if you will in the end have to photoshop it out as a placeholder. Asset generation goes quicker if your base is "okay."

File: Screenshot 2024-08-20 215311.jpg (65 KB, 1201x164)

65 KB JPG

>>101998397
there's this too, it shouldn't have no latents to cache, it's probably why it's crashing
can I see your dataset config? I couldn't find any info about it regarding flux

File: gen_tmp_03.jpg (214 KB, 1312x1312)

214 KB JPG

>>101998453
Reboot vibes

>>101998453
I was going to tell you there's a Chun-Li Flux lora on Civitai but it turns out you have to pay for it

File: ComfyUI_temp_ostyt_00019_.png (1.49 MB, 1024x1024)

1.49 MB PNG

>>101998393
If there was a no text encoder model that was lighter for flux, I would have probably downloaded that one even if it wasn't faster.

>>101998500
this output is already using a y2k lora, I'm not a big fan of stacking up loras, it usually don't work well

>>101998523
>If there was a no text encoder model that was lighter for flux, I would have probably downloaded that one even if it wasn't faster.
????

File: ComfyUI_04990_.png (1.29 MB, 1024x1024)

1.29 MB PNG

https://www.youtube.com/watch?v=SUS64sAksjI

File: 1709112144879620.png (645 KB, 512x512)

645 KB PNG

any suggestions for image upscaling? I am happy to say I finally got it all working on 8gb (and it's the first image gen model that was worth running on 8gb, yay), just gotta upscale the 512x512 to 1024x1024.

>>101998579
This looks very Twisted Metal 3

>>101998594
seems fitting for a y2k lora, it's the same era

>can't run cogvlm2 on windows because of triton dependency which only runs on troonix
why are model devs like this
>inb4 joycaption which hallucinates too much on my current dataset

>>101998672
use WSL

>>101998461
Here's my full command line

accelerate launch  --mixed_precision bf16 --num_cpu_threads_per_process 1 "flux_train_network.py" --pretrained_model_name_or_path "X:/AI/ComfyUI_windows_portable/ComfyUI/models/unet/flux1-dev.safetensors" --clip_l "X:/AI/ComfyUI_windows_portable/ComfyUI/models/clip/clip_l.safetensors" --t5xxl "X:/AI/ComfyUI_windows_portable/ComfyUI/models/clip/t5xxl_fp16.safetensors" --ae "X:/AI/ComfyUI_windows_portable/ComfyUI/models/vae/ae.sft" --cache_latents_to_disk --save_model_as safetensors --sdpa --persistent_data_loader_workers --max_data_loader_n_workers 1 --max_train_epochs 30 --seed 42 --gradient_checkpointing --mixed_precision bf16 --save_precision fp16 --network_module networks.lora_flux --network_dim 8 --network_args "train_blocks=single" --split_mode --optimizer_type adafactor --optimizer_args "relative_step=False" "scale_parameter=False" "warmup_init=False" --learning_rate 0.0004 --network_train_unet_only --cache_text_encoder_outputs --cache_text_encoder_outputs_to_disk --fp8_base --highvram --save_every_n_epochs 1 --output_dir "x:/AI/sd-scripts/outputs" --output_name flux-hinako --sample_every_n_epochs 1 --sample_sampler "euler" --sample_prompts "x:/AI/sd-scripts/outputs/sample/hinako.txt" --timestep_sampling sigmoid --model_prediction_type raw --guidance_scale 1.0 --loss_type l2 --train_data_dir "x:\AI\flux tags ready\work" --caption_extension "txt" --resolution "512,512" --enable_bucket

I had to add

--caption_extension "txt"

because the default extension is "caption" or something, maybe it's because of that?
Also, I remember reading that the learning rate setting is ignored when using adafactor. I saw another anon successfully using AdamW8bit on 12GB, so you'll probably be better off going with it if you don't want to wait for 4000+ steps.

>>101998688
you're a diamond, I'll see if I can do better than 12 hours tomorrow
good hunting

>>101997985
Even 4_K_S takes about 7.6-7.7gb with multiple loras on my rtx 2080. I did need to turn off hardware acceleration in my browser though

>>101998720
>multiple loras
how many loras do you feel Flux can eat before shitting its pants in terms of image quality and prompt adherance?

>>101998743
NTA but 0

File: ComfyUI_00951_.png (1.02 MB, 720x1280)

1.02 MB PNG

>>101997091
>>101997136

File: grid-0434cen.jpg (1.07 MB, 1792x2304)

1.07 MB JPG

Anyone else absolutely hate everything related to Flux is named?
Pro, Dev and Schnell are easy enough, but Schnell is still just as silly as naming it Hayai or fast in some other language.
fp16 and fp8 seem sensible enough at first, but with the next step down it's suddenly nf4.
Then we have Q4_0 which is like nf4 but not quite, but that's a .gguf which I'm supposed to somehow know about, and when I find out I realize you need other files to make it even work.
>ae.safetensors, clip_l.safetensors, t5xxl_fp8_e4m3fn.safetensors (made it work for me, but I have no idea what I'm actually doing)

To make everything more confusing many add even more bullshit numbers an letters.
https://civitai.com/models/647237?modelVersionId=725532
>Flux.1-Dev GGUF Q2.K Q3.KS Q4/Q4.1/Q4.KS Q5/Q5.1/Q5.KS Q6.K Q8
I really have no idea what anything of that means.

While bitching about this thankfully I learned how to use Q4_0 which allowed me to use loras at a sensible speed since I couldn't make them work at nf4 .
That batch took me like 15 minutes at fp8, and being able to do it at a bit over 3 is way more workable.

>>101998743
Each lora messes up the balance, and a badly trained one completely ruins composition and overall quality. I've had 5 loras at most - was curious if realism/amateur photo styles could counter the shitty big tiddy bimbo's bogged faces and they sorta could, but the skin looked even more plastic than without them. It also was quite slow as each lora adds about 10-15% to the generation time.

>>101998824
>I really have no idea what anything of that means.
Q = quant
2/3/4/5 = average bit weight
K = more sophisticated quant method
S = small (because there's a "M" that means Medium)

>>101998824
>That batch took me like 15 minutes at fp8, and being able to do it at a bit over 3 is way more workable.
what gpu, how many steps?

>>101998500
>you vil pay for ze picasso
ah I get it now, saas is the quality $15/month subscription mmo while local is the freemium korean cash shop asset flip.
the absolute state

There are quants of the T5 model now?
https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf
X4KA

>>101998918
>This LoRA will be available for free in 2 days, 2 hours, and 46 minutes or once the donation goal is met.

File: grid-0401.jpg (764 KB, 1792x2304)

764 KB JPG

>>101998879
RTX 3080 (10GB)
25 steps

File: bComfyUI_104851_.jpg (824 KB, 1536x2048)

824 KB JPG

>>101998592
i've seen some people use ultimate sd upscaler but i've never used it. i'm trying to figure out the better method myself.

File: ComfyUI_00952_.png (922 KB, 720x1280)

922 KB PNG

>>101998824
It's all confusing at first but the more time you spend breaking your brain over it the more enjoyable it becomes once you start realizing how it works

File: 1719217501424648.webm (2.33 MB, 1280x720)

2.33 MB WEBM

can i train flux loras with kohya with my 16gb 4080?

>>101999002
Big Space Channel 5 vibe

File: FD_00018_.png (182 KB, 256x384)

182 KB PNG

>>101998251
>>101998266
>You can use Google translate
Я бyдy иcпoльзoвaть cвoй мoзг
>and Photoshop
Heт. The whole point is to do it all in prompt, for me. I don't like to inpaint or photoshop or any of that. It's significantly less about the actual images I produce and almost entirely about the technology around it. I can't imagine being one of the people who spends hours and hours inpainting and fine tuning and photoshopping an AI output. At that point you may as well just draw it yourself.

whats a site to generate songs?

>>101999019
indeed, that's what I was aiming for

>>101999038
suno.ai udio.com
also wrong thread, retard

File: 00102-2466508696.png (1.81 MB, 1024x1440)

1.81 MB PNG

>>101999011
I've been having great luck with this config, using bmaltais gui fork
https://github.com/bmaltais/kohya_ss/issues/2701#issuecomment-2297761417

File: 1719532311706921.png (1.23 MB, 1212x785)

1.23 MB PNG

>>101998990
Ended up using a general purpose ESRGAN compression remover called "2x Pooh V4". Seems to work alright, dunno if it's the "BEST" but it lets me have 1024x1024 under 8gb and I'm satisfied.

>>101999073
thanks a shit ton. how long does it take compared to sdxl?

File: 1723858003283116.jpg (128 KB, 825x850)

128 KB JPG

>>101998824
Both Q4_K_M and Q4_K_S provide more accurate results than Q4_0 or Q4_1 (at least with language models), so use one of them.

File: file.png (2.51 MB, 1024x1024)

2.51 MB PNG

>>101998824
>. venv/bin/activate
>gen
simple as

In all seriousness, once you spend a few days with Comfy everything just clicks and it's super enjoyable. But use a model and settings that don't make it a slog. At 15 min per gen you cannot enjoy yourself.
>>101998973
Use nvidia-smi to see how much VRAM is being used and use a quant that leaves enough space so that the whole thing happens in GPU. Make sure Comfy says "loaded completely".

File: 1724189802.png (928 KB, 1024x768)

928 KB PNG

>>101998824

File: ComfyUI_00953_.png (886 KB, 720x1280)

886 KB PNG

JoyCaption seems to break down after each run, it works fine in the first run and after but if I change the picture I want to caption a new it breaks for whatever reason, anyone knows why?

File: file.png (1.89 MB, 1024x1024)

1.89 MB PNG

>>101999082
Nice
>>101999095
But keep in mind, Q4_0 is faster than any of them.

File: 00135-2863446420.png (1.26 MB, 1024x1024)

1.26 MB PNG

>>101999086
I never trained SDXL loras.
It took around 1 hour 40 minutes. I changed that config from 1 epoch 1000 steps to 15 epochs 3000 steps though. Trained on 20 images

Can an anon familiar with how Flux/SD works explain the flow of basically commands to the gpu?

Like my understanding is the cores/streams are basically shaders, ie each runs what? the same sorta-program? But what determines those commands?

File: file.png (2.63 MB, 1024x1024)

2.63 MB PNG

I've been noticing Flux ignoring my prompting for text in the image. Is this a sampler/scheduler thing?

File: ComfyUI_00955_.png (879 KB, 720x1280)

879 KB PNG

>>101999140

>>101999184
>Trained on 20 images
i've always wondered, is 20 images enough for it to learn both the body and the face?i've trained sdxl and ponyxl loras with 100 images before. i'll try 30 for flux to start with i guess.

>>101999140
Does it give you any errors? Or simply refuse to caption the picture? If it only refuses, try switching to a different picture, captioning it, and then switching back, this time it will work.

>>101999095
fp16 is actually better? idk, I get a slight boost from q8, not sure if I should use it. I am rn, and don't see any real issues...

>>101999058
[spoiler]Is there anything non-shitty that's local in audio? like an instrument that follows a midi score?[/spoiler]

>>101999224
fp16 is the baseline, the foundation, the holy grail. All quantizations are judged based on their closeness to fp16 results.

>>101999204
to get better consistency on that you have 3 solutions:
- Transform your prompt into slop boomer prompting with chatgpt
- Increase your CFG (cfg 6 + Tonemap works well)
- Increase your number of steps, sometimes it makes the text appear

File: 00010-4253160097.png (1.97 MB, 1024x1440)

1.97 MB PNG

>>101999209
Seems like it, it can do expressions and different angles well. Face can be a little blurred in distant shots sometimes though. Should've included more full body photos in the dataset I guess

File: 1742-flux1-dev-1229769286.png (1.53 MB, 1024x1024)

1.53 MB PNG

The fucking bimbo look is killing Flux for me. I don't want to generate boring airbrushed basic bitches, I want cute women with some flaws and unconventional beauty. Even getting images like this with conventionally beautiful women with different phenotypes is like pulling teeth.

>>101999284
looks extremely good. tags or no tags?

>>101999292
have you tried the realistic lora? or the "boring snapshot circa 2015" prompt?

File: tod_.png (2.85 MB, 1336x1336)

2.85 MB PNG

>>101999292
i have to edit every flux gen with sdxl, wish I had a the real version

>>101999243
Audiocraft maybe?
https://github.com/facebookresearch/audiocraft
>amdlets need not apply

File: ComfyUI_00954_.png (764 KB, 720x1280)

764 KB PNG

>>101999220
I am getting this error

Error occurred when executing Joy_caption:
Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules in 32-bit, you need to set `load_in_8bit_fp32_cpu_offload=True` and pass a custom `device_map` to `from_pretrained`. Check https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu for more details. 
  File "ComfyUI\ComfyUI\execution.py", line 316, in execute
    output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
  File "ComfyUI\ComfyUI\execution.py", line 191, in get_output_data
    return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
  File "ComfyUI\ComfyUI\execution.py", line 168, in _map_node_over_list
    process_inputs(input_dict, i)
  File "ComfyUI\ComfyUI\execution.py", line 157, in process_inputs
    results.append(getattr(obj, func)(**inputs))
  File "ComfyUI\ComfyUI\custom_nodes\Comfyui_CXH_joy_caption\Joy_caption_node.py", line 146, in gen
    joy_pipeline.parent.loadCheckPoint()
  File "ComfyUI\ComfyUI\custom_nodes\Comfyui_CXH_joy_caption\Joy_caption_node.py", line 94, in loadCheckPoint
    text_model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, device_map="auto",trust_remote_code=True,quantization_config=quantization_config) #
  File "ComfyUI\python_embeded\Lib\site-packages\transformers\models\auto\auto_factory.py", line 564, in from_pretrained
    return model_class.from_pretrained(
  File "ComfyUI\python_embeded\Lib\site-packages\transformers\modeling_utils.py", line 3865, in from_pretrained
    hf_quantizer.validate_environment(device_map=device_map)

>>101999304
I don't actually want realistic images though, I want something like the style of that image I posted. Maybe I can mix it with another style LoRA.

File: ComfyUI_00956_.png (786 KB, 720x1280)

786 KB PNG

>>101999220
And no, that doesn't work, once I switch a pic the error persists until I restart comfyui

>>101999340
How about this with a strengh of 0.3/0.4, it would stop the bimbo effect
https://civitai.com/models/635218/flux-dev-ugly-things

>>101998579
wtf is this meme

>>101999284
can it do puckered lips?

File: 00019-1495168978.png (2.46 MB, 1024x1440)

2.46 MB PNG

>>101999302
I experimented a bit.
So far I've found just a short sentence describing the scene works best, something like
>Side profile photo of Elizabeth Olsen with short blonde hair wearing a maroon velvet blouse and earrings in front of some purple flowers
I tried full natural language boomer captioning and booru tags and got worse results. I'm no expert though

File: file.png (2.64 MB, 1024x1024)

2.64 MB PNG

>>101999265
I can increase guidance too, but the image looks less interesting. And the speech bubble is not on the person it's supposed to no matter what I do.

>>101999292
This looks very nice. And I feel the same way.

File: ComfyUI_05002_.png (1.25 MB, 1024x1024)

1.25 MB PNG

https://youtu.be/HviBn_L6quA?t=33

>>101999385
>>Side profile photo of Elizabeth Olsen with short blonde hair wearing a maroon velvet blouse and earrings in front of some purple flowers
oh damn, yeah all my training data is,captioned,like,this. i hate this flowing text meme. i probably have to recaption everything then

>>101999073
this is amazing, only the grill meshwork and a couple things on the toes and wheels give it away. Especially impressed with the iron posts staying true and straight through the back window.

>>101999401
>I can increase guidance too, but the image looks less interesting.
that's why I increase the CFG instead, it helps for the prompt adherance without ruining the style

>AssertionError: fp8_base requires mixed precision='fp16' or 'bf16'

but it doesn't accept

--mixed_precision bf16
--mixed_precision 'bf16'
--mixed_precision "bf16"
--mixed_precision=bf16
--mixed_precision='bf16'
--mixed_precision="bf16"

thank you kohya very cool

>>101999417
the holes of the pants are wrong and one leg is longer than the other too

File: ComfyUI_05003_.png (1.44 MB, 1024x1024)

1.44 MB PNG

https://www.youtube.com/watch?v=fiCRhszDHlM

>>101999446
yeah, but you never know when some weird hollywood clothing designer is going to make pants where both legs come out one hole. Or maybe it's a new illuminati sign to have secret pants holes for your feet so you can trick people into thinking real pictures are ai or something.

File: file.png (2.39 MB, 1024x1024)

2.39 MB PNG

>own

>>101997581
nf4 is faster than Q4_0 and generates slightly higher quality in my personal experience
i have no idea why everyone keeps shitting on it, i hope its not because of that one miku skateboard image because thats retarded

>>101999459
why do you people know all these weird ass memes

File: ComfyUI_temp_vukzf_00077_.png (1.41 MB, 1240x792)

1.41 MB PNG

first actually decent painting style lora for Flux dropped
https://civitai.com/models/661394/eldritch-oil-painting-or-for-flux1-dev

testing it now, it's the first one I've found that strongly imparts a painterly style and removes the slop look without also breaking the model's coherence and making it retarded

>>101999503
>i have no idea why everyone keeps shitting on it,
even on the LLM ecosystem no one use nf4 and prefer to use Q4_0

File: 1803-flux1-dev-1720095810.png (1.6 MB, 1024x1024)

1.6 MB PNG

>>101999359
Good suggestion. I'm not where I want to be but I think I'm getting somewhere.

>>101999525
NTA but you didn't answer the (implied) question. Why? Other than "because everyone else says so".

>>101999536
The same question could be asked to you, why do you prefer nf4 to Q4_0? Do you have any tangible proof it's better?

>>101999542
I use Q8_0. And I use it because it's the largest quant I can use at full speed.

>>101999542
See, you inferred that I'm saying NF4 is better. I'm not. I have no idea either way. What I'm asking is why more people are using Q4_0. That's all.
Any antagonism you sensed is in your head.

Is there a way I can extract the prompt and seed from the metadata of an image in ComfyUI?
I have an upscaler workflow and I want to pull that data into it.
Currently I am loading a duplicate workspace and manually copy pasting it.
With other models I would just upscale every image but Flux is too hungry.

File: file.png (599 KB, 512x512)

599 KB PNG

>can almost gen an SD1 image per second now
haha take that past frustration

>>101999572
>See, you inferred that I'm saying NF4 is better. I'm not.
>>101999503
>nf4 is faster than Q4_0 and generates slightly higher quality in my personal experience

File: ComfyUI_05004_.png (1.27 MB, 1024x1024)

1.27 MB PNG

>>101999498

>>101999605
I said here >>101999536 "NTA", which I assume means "not that anon". I hope it doesn't mean "not to argue" because that means I've been an absolute newfag for almost a year now using it wrong.
tldr you quoted two different people

>>101996391
https://discord.com/invite/Y4aH5KubP8
4chan /ai/ server

>>101999587
Prompt Extractor from Inspire Pack or Metadata extractor from Crystools

File: ComfyUI_00961_.png (958 KB, 720x1280)

958 KB PNG

>>101999327

>>101999626
Thank you, Anon

File: file.png (2.12 MB, 1024x1024)

2.12 MB PNG

>God says "whatever dude"

File: FLUX_00044_.png (1.57 MB, 896x1152)

1.57 MB PNG

kek, this ugly lora works great

>>101999623
but the other anon is saying that nf4 is better than Q4_0, are you also gonna ask him why he feels that way or are you only gonna interrogate me?

>>101999645
That's a man...

>>101999647
... sigh
That anon compared the two and said "generates slightly higher quality in my personal experience". Meaning, he likes the results better. I don't need to "question" him. I just want to know why different people have different opinions on NF4/Q4 so that I can learn something about this tech, hopefully.
I'm going to go install NVIDIA drivers on my Windows partition. Have a great rest of the day.

>>101999674
I was gonna say this, looks like a twink with a wigh

File: st,small,507x507-pad,600x(...).jpg (37 KB, 600x600)

37 KB JPG

>>101999680
>... sigh

>>101999625
What the fuck was that place

File: Screenshot 2024-08-21 102304.png (44 KB, 885x448)

44 KB PNG

>>101999626
Hmm no seed or am I missing something? I suppose I can use a random noise for the upscale, it's low denoise

>>101999725
Just report and move on.

>>101999725
Alice in Wonderland, but with Alice fully naked and being done in by unspeakable things

>>101999783
Context?

File: ComfyUI_00965_.png (883 KB, 720x1280)

883 KB PNG

>>101999629

>>101999647
i just feel that way, really
just from using both models and comparing them on the same seeds, i could often get better stuff out of nf4 when giving it the right amount of steps
the only drawback i experience is that it's a bit more volatile around 20 steps, and changing it by 1 in that range can change the image quite a bit, but that also increases the odds of getting a lucky draw
this is all just from personal experimenting cause i refuse to trust a single benchmark thats not my own eyes
but yeah if i can claim one thing confidently its that its not "shit" compared to q4 and anyone who calls it that is probably just parroting other anons who also never ran it
so yeah theres my answer
>>101999623
no you're good, ive never seen nta get used for "not to argue"

File: ComfyUI_05007_.png (1.24 MB, 1024x1024)

1.24 MB PNG

I miss the old Idubbbz
https://www.youtube.com/watch?v=P-_GWUw8LwM

>>101999184
Are you training locally? How much VRAM?

has anyone been able to run ViT-L-14-BEST on forge? I can't seem to get it to work
also wondering if it would actually make a big difference or not

File: ComfyUI_05008_.png (1.39 MB, 1024x1024)

1.39 MB PNG

https://www.youtube.com/watch?v=HwJcNCWrVD0

>>102000021
It does make a difference on Comfy, so I can answer that part of it. No idea how to make it work on Forge. Are you up to date?

File: ComfyUI_00973_.png (844 KB, 720x1280)

844 KB PNG

>>102000011

>>102000047
yes and ive tried both the safetensors and the statedict file but it always throws an error about a missing clip state dict regardless

File: ComfyUI_00974_.png (870 KB, 720x1280)

870 KB PNG

>>102000071
>>102000000
>>101999999

>>102000076
Well that's all the advice I can give. Sorry Anon, hope you figure it out.

>>102000099
wasted as usual

>>102000099
Epic win for comfy

>>101999806
That's super cool, what was the prompt? Looks sort-of like the images you get from an electron microscrope

>>101999942
More

1. You can train a lora of yourself
2. You can perfectly blend anime girls into real photos
3. ???????

>>101999344
Watch your ram utilization, in a process manager.

>>102000206
4. It won't fix your loneliness

>>102000206
????? indeed because your plan well you see it sucks

>>102000030
>https://www.youtube.com/watch?v=HwJcNCWrVD0
basado

File: ComfyUI_00977_.jpg (153 KB, 720x568)

153 KB JPG

>>102000239
I still have room for another 15GB or so, the whole LLM and other stuff should be around 9GB, not to mention how it runs fine at first which I take would mean it's already loaded in the ram.
My own guess is that it's trying to load the models back to the GPU but can't find enough vram because Q6 keeps 7GB there after the run.

>>101999975
Yes
16gb

>>102000305
Did you WATCH IT?

>>102000325
Yes, I am always watching in case something goes wrong

>>102000318
I'm going to try it out :^)

>>102000305
If windows has something like nvtop you can tell if a process is freeing VRAM or not. If it doesn't have something like nvtop find something windows default shit is trash.

File: bComfyUI_105140_.jpg (314 KB, 768x1024)

314 KB JPG

>>102000145
think it was a mix of random shit like spumellaria, quasicrystal, self-similar, Koch snowflake, phyllotaxis.
i'm not boomer prompting yet just seeing what i can do with flux using minimal words.

does someone know what's the maximum prompt length for flux?

>>102000381
You stupid cunt, I had an le epic image lined up for the final image and you wasted it with this shit.

>>102000347
Wait I will post the screenshot in the new bread

>>101999265
>Tonemap
What's that?

>>102000394
not my problem

baked, one min

>>102000423
it's a node that allows you to go for CFG > 1 without burning the picture on flux, you can find its node here
https://github.com/comfyanonymous/ComfyUI_experiments

Praying for a good collage

>>102000437
btw, is there a list of ways to generate weird effects? Burning is one, but idk what all of them are. I have also found using the wrong vae decoder can be neat. also, image sizes too small, especially with too few steps (with sd, I guess I need to try it on Flux too).

(for a glitch effect aesthetic)

anon?

>>102000428
why do you build me up (build me up)
buttercup baby, just to let me down (let me down)

>image cap
>post cap
it's over

>>102000567
owari da

>Error running job: CUDA out of memory. Tried to allocate 314.00 MiB. GPU 0 has a total capacity of 24.00 GiB of which 0 bytes is free. Of the allocated memory 21.63 GiB is allocated by PyTorch, and 1.59 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

suck my fucking dick

ETA: 200s

>>102000655
>>102000655
Just do what it says, set pytorch thingy. I set thingy too, for amd, and the amd thingy works. Like you have to do it every time in the shell, idk how it is on Windows.

You set a variable that is telling ***CUDA*** drivers what to do. (in my case amd rocm drivers)

Fresh bread from the oven
>>102000715
>>102000715
>>102000715
>>102000715
>>102000715
jump in