LORA integration
Quick question, your new update for 5.2 mentions that you "Updated "snofs" and "qwen4play" NSFW LORAs + Meta4 for v5.2. " Does that mean the lora are already in your checkpoint, or I still need to load one of those loras and you are just saying its compatible with them?
Pls! Can you add to the beginner's guide how to use loras and other setting for model. I get unrealistic and uninteresting images for nsfw assembly. I can't find the information on the websites.
I also came across the fact that any level of denoise makes the image with img2img worse. euler_a/sgm_uniform all body looking plastic with Professional digital photography"(
What could be the problem?
@Ok333y , I know what you mean, euler usually produces plastic-y skin. I found the most realism using lcm/ddim_uniform. Give that a try. Makes the images a bit darker as far as lighting goes, but the realism couldn't be matched by the others.
Quick question, your new update for 5.2 mentions that you "Updated "snofs" and "qwen4play" NSFW LORAs + Meta4 for v5.2. " Does that mean the lora are already in your checkpoint, or I still need to load one of those loras and you are just saying its compatible with them?
Yes, those LORAs are already mixed in the "all in one" and don't need to be added after.
Avoiding plastic skin is kinda tricky, especially at higher step counts. Qwen likes its plastic skin.
Penises are pretty bad when prompting, any tips on how to get that looking better? Any good loras to recommend for genitalia? I was using a qwen penislora, but even that isn't very realistic.
@Ok333y , I know what you mean, euler usually produces plastic-y skin. I found the most realism using lcm/ddim_uniform. Give that a try. Makes the images a bit darker as far as lighting goes, but the realism couldn't be matched by the others.
how many steps and cfg and denoise level?
Quick question, your new update for 5.2 mentions that you "Updated "snofs" and "qwen4play" NSFW LORAs + Meta4 for v5.2. " Does that mean the lora are already in your checkpoint, or I still need to load one of those loras and you are just saying its compatible with them?
Yes, those LORAs are already mixed in the "all in one" and don't need to be added after.
Avoiding plastic skin is kinda tricky, especially at higher step counts. Qwen likes its plastic skin.
with a small number of steps, the image is blurry and poor, at steps up to 35 it is so plastic, above that the plastic is already flowing. Do you have any ideas?
I just use the workflow from the OP. Takes me about 50 or so image creations for it to get the faces and the pose right. Once I get a good shot, I just send that through an upscaler to cleanup the image. Realism of the faces is most important, as long as the pose is close I can animate it to my liking.
I just use the workflow from the OP. Takes me about 50 or so image creations for it to get the faces and the pose right. Once I get a good shot, I just send that through an upscaler to cleanup the image. Realism of the faces is most important, as long as the pose is close I can animate it to my liking.
img2img?
Yes, img2img. Its just so hard getting realism when you are feeding it your own pics/characters. Might have better luck with a character lora, but those take ages to train locally unless you want to pay to have them done at runpod or somewhere like that. I literally just let the workflow run like 50 iterations for whatever pose I want the characters in, I get maybe 1-2 normally that match what I asked for and the faces are really close to the original.
Yes, img2img. Its just so hard getting realism when you are feeding it your own pics/characters. Might have better luck with a character lora, but those take ages to train locally unless you want to pay to have them done at runpod or somewhere like that. I literally just let the workflow run like 50 iterations for whatever pose I want the characters in, I get maybe 1-2 normally that match what I asked for and the faces are really close to the original.
Thanks for the answer!
You're really helping the newbie. Can you show me your workflow and settings?
I've roughly generated 500-600 images on different samplers, and all of them aren't even close to the original face.
I tested your story with lcm/ddim_uniform and the results were even more plastic for me(
you could give https://civitai.com/models/1939453/qwenedit-consistence-lora a try, maybe a future v5.3 addon, just ggufing v5.2 an then if done ill try loading it with "Power Lora Loader (rgthree)" additlionally.
weight ~0.48 for single image.
@phil2sat I will give that one a try and see what my results are. Also, are you planning on integrating an 8 step lightning/light2v? I have a 4070 and usually do all my generations at 8 step for the quality.
@Ok333y , I am still playing with the denoise value and the amount of steps to get some consistency. Let me do some more testing and I will share with you where I think the sweet spot is. Also going to test that lora Phil mentioned above to see if I can get closer to getting away from plastic skin. Its a constant battle to try and find the right values.
@phil2sat I will give that one a try and see what my results are. Also, are you planning on integrating an 8 step lightning/light2v? I have a 4070 and usually do all my generations at 8 step for the quality.
this is not my work, i just use it and upload gguf files to huggingface out of this repo.
but this are the actual loras for v5.2, this is a testing workflow of mine only missing the lightning fp8 lora.
to check whats inside just click on the saftensors file and there is a prompt and workflow section where you can see what loras used.
for some case im unable to reproduce them my system ram is going up like hell and the output doesnt give same results as if i use the original for my ggufs.
Used the qwen2.5 model to self produce it a preservation prompt, this is the result and it doesnt look that bad:
Preserve the person's unique visual identity by maintaining their face structure, expression style, proportions, skin tone, and body build. Allow for creative modifications to other elements such as hairstyle, outfit, background, and lighting.
maintain camera angle.
but to be honest, im trying hard to get exact the same person, tried sdxl, pony, flux, qwen, qwen 2509, i get close to but not the same.
actually testing the lowest possible quant Q2_K but even if it gets loaded completely it takes around 45-100s/it making testing really time consuming.
also i guess it needs some more steps to get rid of the grid, but not that bad
@phil2sat From the above breakdown, it looks like "snofs" is best within your checkpoint. How would I isolate just referencing that in my prompt for NSFW generation. I am getting closer to getting the realism and close to the original characters as I can, but it still takes multiple generations to get it close to what I want, that is just a function of qwen. Wan is much more consistent with results, but less so for prompt adherence. I am hopeful Wan 2.5 will make this easier/better. I will post my results when I am done testing. Great work though, your checkpoint is literally the only one that worked out of the box with your workflow for generating NSFW images consistently, just having to play with the scheduler/sampler/denoise values. Oh and in the ksampler, 8 steps is king, the quality is a lot better. Once I get the denoise right I will play with the steps to find the sweet spot, then try some lora integration to remove the plasticy-ness
As i told thats not my repo here.... i do only the GGUF files from Phr00t's safetensors
thats what i use https://huggingface.co/Phil2Sat/Qwen-Image-Edit-Rapid-AIO-GGUF cause i cant run the large workflow, stupid stoneage AMD-GPU upcasts everything to FP32 and OOM's alot.
and while i wait that my stoneage pc finishes its work i have a lot time to write, lol
normally i go to the civitai lora page https://civitai.com/models/1972981/qwen-sex-nudes-other-fun-stuff-snofs
and check prompts they used for the images posted.
Noted, I will give yours a try, working off of Phr00t's workflow. Just trying to see what combo works best :) I appreciate everyone's work.
btw in the image above theres also my workflow included, but actually i only have v5 online:
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1933/1933 [12:37<00:00, 2.55it/s]
INFO:gguf.gguf_writer:Writing the following files:
INFO:gguf.gguf_writer:/daten/models/Qwen-rapid/tmp/qwen-rapid-nsfw-v5.2-F16.gguf: n_tensors = 1933, total_size = 40.9G
Writing: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40.9G/40.9G [20:01<00:00, 34.0Mbyte/s]
waiting, waiting, waiting, testing, waiting....
OK I spent literally 6 hrs testing a million combinations, and I think I found a good mix of realism.
I am using the new 5.2 checkpoint, using the json provided by the OP, using the following settings in the ksampler:
I am also using this lora to cleanup the skin texture a bit to be a little less plastic: https://huggingface.co/olesheva/real_life_lora_qwen (run this at 0.4)
Next, going to try @Phil2Sat version as well to see if that nets any better results. You will still have to do a TON of iterations to make sure you get a good result with prompt adherance, but the results look good right now. Hope this helps.
Easy way to generate different results, set the "control after generate" to INCREMENT and then just queue up a bunch of runs, it will add 1 to the seed and give you all different iterations of your prompt.