what browser you use to navigate? i use Microsoft edge.

@plalsaga70717 Tried with Chrome, Firefox and Edge. On Chrome and Edge it downloads but then stops and says network error.

@zfreakazoidz496 same for me, but i retry and it work,refresh the download button

@KandooAI Error(s) in loading state_dict for CLIPTextModel:

Missing key(s) in state_dict: "text_model.embeddings.position_ids".

same

It seems to be a vae issue. Try bypassing the stable cascade's vae and using the baked in one instead.

Are you sure you are using the suggested cfg of 1?

Even with (((realistic, photograph))) in the negative prompt the results are still too photographic. Well, the V9 checkpoint has "photo" in its name so I guess I can't complain too much, but the outputs are just so different from V7, which is what I've been using up to now. Thankfully I have a recent graphics card so it only takes around 20 seconds per image with V7.

@Afgmc It is the low steps and low CFG that cause most of the prompt words to fail

@wyxzddsjj919 Isn't low steps the whole point of lightning models? Anyway, like I said it's no problem for me to stick with the "snail" models, I just hope the whole ecosystem doesn't switch over to lightning.

@wyxzddsjj919 I think you got the point. All turbo/lightning models use 1~2 CFG, which basically ignores prompts. However, if you want less steps you cannot use high CFG since the result will be total crap.

We’re working on something exciting. Stay tuned.

@KandooAI Thank you for your thorough answer. Appreciate it <3

@KandooAI I am seeing a trend on lightning models to overly smooth human images. Skin texture has all but disappeared.

@KandooAI Thank you I will

@KandooAI what about overall ... creativity lets call it. I've tried a few lightning model variants, and also adding a LoRA even on top of the v6, and despite the fact I can still see the same underlying composition the diversity of results is reduced.

Not sure if it requires a change in prompting or it's the price one pays for distilling the models

upper body analog photo of beautiful slavic girl, white hair, gigantic breasts size, blue coat, white blouse with deep cleavage

Negative prompt: CGI, Unreal, Airbrushed, Digital,

Steps: 5, Sampler: DPM++ SDE Karras, CFG scale: 2, Seed: 3385708707, Size: 832x1216, Model hash: c8df560d29, Model: juggernautXL_v9Rdphoto2Lightning, Version: v1.7.0

I dunno, 3 times generated and girl was clothed :D

(fully clothed:1.3) and put it closer to the start of your prompt

1.0? You mean the full version?

Lightning is faster. Good quality. You’ll always get better quality with the full version v9

I loved V7

Lower CFG

Negative prompts “plastic”

Reinforced Positive prompts like “Detailed skin” help too

@RunDiffusion thank you. Will try.

V7 was the best version, the more a new version is released, the less realistic the output is

@valer_85354 is there a way to use v7 to generate? v8 is not available to generate, and so I downloaded the v8 safetensors file and added a new model, but that is also not available to generate. Any advice?

@Amateur007 no idea, I was wondering the same. but it's up to civitai to make it available to use here

RD Photo V1 and V2 anew available on our platform for a really inexpensive price. Need to keep some things behind closed doors. It’s the only way to fund this research with KandooAi. Maybe when we get access to SD3 we’ll open source all the SDXL models. Who knows! As always we appreciate your support!

You can run on ComfyUI. I use it.

@gokhanh3333 I have a workflow that works with other realistic models but when I use this I get Error occurred when executing VAEDecode. any clue why?

I tried with A1111 and seem like im having some memory issues : "OutOfMemoryError: CUDA out of memory. Tried to allocate 40.00 MiB. GPU 0 has a total capacty of 8.00 GiB of which 0 bytes is free. Of the allocated memory 6.83 GiB is allocated by PyTorch, and 476.19 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF"

@J19Z A1111 is bad in handling memory, Try use sdwebui forge, there is an integrated extention called "neverOOM" that can save vram, so Out Of Memory error would not happen frequently. Enable the extention before generate an img.

Assuming you are using "juggernautXL_v9Rdphoto2Lightning.safetensors", try steps: 5, cfg: 2, sampler: dpmpp_sde, schedular: karras. Use a latent of width: 832 and height: 1216. This produces great results for me, but do experiment with steps and cfg. Also, depends on what you are trying to make, not all checkpoints are suited for all things, though this one is pretty robust.

@KandooAI i agree, so what are you going to do next then, just wait for sd3?

@KandooAI ok good idea. look up the proteus model, its pretty good as well. i think it was captioned by gpt :D

controlnet

are you using SD1.5 LoRAs or embeddings by any chance? this looks pretty fucked up. or perhaps your VAE? try setting it to "automatic"

nope, at least not that i know of. are there different SDXL VAE's?

actually just realized i don't know where some of the recommended settings are? Sampler: Siax_200k

Steps: 2

Denoise: 0.35

Upscale: 1.5

ok found the stuff under hires fix, but i can't find where that sampler is set...it's not in any of the dropdowns i can find. Sampler: Siax_200k?

@rastoboy for the image use 4- 6 steps, in the hires fix tab use 2 steps for upscale. images should come out better.

If you put (hat) in negative prompt, it should remove hats. What happens if you put (no hat) in negative prompts? Does the double-negative come out as a positive? Is that what you're doing?

Be sure that you are making affirmations only on the negative prompt. If you write (hat), you are basically saying "I don't want hats in this picture". If you write (no hat), you are saying "I don't want a picture without a hat". Also, a good technique is to write with hyphens, it makes what you write explicit. A prompt saying "a-woman-wearing-a-hat" is explicitly sayin "I want a woman wearing a hat, do not do anything other than that". Hope this helps!

Personally I prefer the image compositions produced by 3M samplers but I think the colors are more consistent with 2M, less artifacts.

Why DPM++ 3M SDE Exponential?

@Bra2ha On paper everyone has all the requirements to be superior to others

@anelesmelchiorre Can you elaborate?

In Automatic1111, I believe DPM++ 2M Karras is the default. Probably why so many use it, by default. ;) For me, I tried a bunch of them on image gens I liked, then judged on that. For me, I actually do like the default best. If I switch to DPM++ 3M SDE Exponential now, I'm so used to DPM++ 2M Karras that all the faces and stuff just look a bit "off", so back to DPM++ 2M Karras. :D

tl;dr: Use whatever you think looks best for your art and don't worry about others.

Run an xy plot and see what you like. It's all up to so many different factors, model, cfg, settings.. turbo, lightning, 1.5, sdxl.. etc..

@anelesmelchiorre Also on paper, everyone has the requirements to sound sarcastic and not be taken serious, like your self.

Strong negative prompts?

@davy272 I've tried negative prompts of all kinds, but I couldn't get it to make like I want it without creating distortions or the model simply ignoring me.

@RheisSivvi  Which lightning model? You're using it with Juggernaut? (XL?) Or instead of?
What is it supposed to bring to an image? Is it simply just default lightning in images, or something more... structural, ie a model, not a LoRA?

I've never found Juggernaut XL to be particularly lewd. In fact, I had to work hard to prevent it being extremely tame. (reverse of your issue, I know).

Distortions and/or ignoring of a prompt always makes me check the CFG and steps, and that's almost always the issue. Only a couple of times has a LoRA caused it to go haywire.
Sorry I can't be much help.

Yes. It's definitely harder and harder to get anything other than photorealistic since then. It's sad because it was so good overall model not just photo one.

tell more please..

I get that occasionally too, with a variety of models. Solo and multiple or 2 in the positive/negative as appropriate - and a fair bit of just rolling again. :-)

If it helps, I just posted some images in 9:16(https://civitai.com/posts/1840089). I had no issues...but also didn't use any of what I said previously, so maybe ignore that, hah. My prompts are in the post, rather than the image - maybe they'll be of use.

Not sure if you're talking about ethnic diversity or not, That's a great thing and something most checkpoints lack a lot. That aside, I much prefer version 8 in my testing.

Use real life or 3d models images as reference for image prompt and PyraCanny.

oh sorry i probably should have mentioned i'm doing txt2image

i do see that it works ok with very short prompts....but not longer ones.

@rastoboy it's fine even with longer ones, it all depends on what your trying achieve. https://github.com/lllyasviel/Fooocus/discussions/557

@blacis or i guess what you're saying is that's a better technique. google is amazingly short of information on what PyraCanny is...gotta link by any chance?

@rastoboy I recommend using the model in comfy, and implement openpose + canny + zoe depth map control nets for an image you want to replicate. SD in general is still open source can be rigid sometimes when it comes to regular prompting.

@KandooAI hope a lightning model for us with old hardware, also finding models are single character is hard to do a (cat behind a mouse) seem to get merge creature

me to

the same any solution ?

im still not sure with sdxl the exact way to do this but if you're looking for better racial representation adding noise to the image helps. its generally how to get things darker. how to get that isn't so straightforward i'm on the path to figuring out the same thing but it was pretty easy in 1.5. might need to go further and composite characters generated with seperate weights on the noise from their environment if it's light out though as doing this tends to affect things like shadows