You can make your own in checkpoint merger. In primary model use sd-v1-5-inpainting, in secondary model use this (realdream) and in tertiary model (c) use sd-v1-5. Select "add differences" in interpolation mode, set the multiplier to 1 and merge. Can grab the sd-v1-5-inpainting and sd-v1-5 checkpoints on huggingface or anywhere else on the internet.

@sinatra oh, thanks for clarification. Keep it up :)

You're absolutely correct. While I don't personally reject NSFW images, especially when they exhibit high-quality and artistic content, I still appreciate them. However, as we've observed, the gallery is inundated with low-quality and repetitive images, and unfortunately, there seems to be little we can do about it. Quite disappointing indeed.

@BIG_A i strongly wish people people upload nsfw (even sfw) with artistic perception rather than being completely pervert.

i (we all) prefer quality over quantity.

it helps to stay focused and achieve better prompts

@sinatra but your model is a merge, not trained/finetuned, no? It's difficult to improve just by merging, but it can improve with more training

@sinatra It would look like a photo instead of CG. More like this: https://civitai.com/images/2466364

Note: Possible errors because it is a translation from Spanish (Spain) to English.

XD I do not know.

Try to copy the exact prompt from an image of the model.

Another thing you can do is little by little.

Try a tag and try,

then add another and try

Until you find the tag that generates that.

Finally, I can leave you a simple prompt that I usually use as a template.

(I have omitted the NSFW that I usually put) ;)

Note: You can replace "german woman" with whatever you want. I generally like it and I usually change it for the name of some lora.

▼ Prompt positive

"german woman"

in love with you

,sunny day, hard lit, accurate her "body measurements" of her, character with details, accessories

▼ Prompt Negative

EasyNegative, censored, extra arms, extra legs, disfigured face

───

Important: The "EasyNegative" tag is a help embedding.

Link if you want: https://civitai.com/models/7808/easynegative

I'm recently using a lora which improves the details.

It's called "Add More Details - Detail Enhancer / Tweaker (细节调整) LoRA"

https://civitai.com/models/82098/add-more-details-detail-enhancer-tweaker-lora

I use it with a power of 0.5

▼ This is my base configuration.

I don't use: Restore faces, Tiling or Hires. fix

Width: 512

Height: 616 or 720

Batch count: 1

Batch size: 1

CFG Scale: 4

───

I hope it helps.

You can use the add detail lora with minus weighting on the other models to achieve the same output :-)

@AndromedaPerson I wonder if there was a misunderstanding in what I wrote. I am already extremely pleased with the minimalist production of this model. Many other models cannot do this. It creates unnecessary objects and unnecessary details. I wonder if I could have expressed it better?

Agreed, I appreciate the 'minimalist approach' too. It leaves room for detailed descriptions, Lora's and some advanced prompting, without losing its versatility.

@sinatra I am using your models as refiner usually. But i will test them as base model. And tell you which one is my favorite.

kubilayan, when you were making your original comment, which version of Real Dream were you referring to? And has your opinion changed as newer versions came out? Which is your favorite version of Real Dream and why?

@llmaps I downloaded and installed V10 yesterday. And I used it. I still think it's a very good model. My opinion hasn't changed. I have previously prepared comparison images and presented them here. I didn't have much time after that. Now I will do a detailed study again and share it with you. I definitely make this right before or after Christmas. I can already say that my main idea has not changed. I also use other SD 1.5 models. And I generally use these as refiners after SDXL models. This is my favorite model.

@kubilayan Thank you very much for answering my question!

I've developed a theory that many of the top models can be different but equal between versions, and have collected multiple versions of the top 15, of which this is one. I have currently retained 7 and 9 upon testing, and if 10 makes the cut I will keep that too. Even Cyberrealistic with it's many, many versions only had 3 that I considered top quality, and they weren't the latest 3, although the most recent versions of most models have tended to be keepers. All this to say, regardless of what the poster thinks, you should test and draw your own conclusions to make a decision based on your own preferences.

I went on this journey being initially disappointed in some latest versions, to see once and for all if newer is really always better, which I don't think it is. My testing has been based on the finer details of realism, like skin detail and eyes, as well as some compositional factors like overall style, diversity (a model making more than one face or body type between seeds for the same prompt), with hands and feet winning bonus points. This is likely all intuitive for you, as it was for me, but I definitively proved it for myself out of curiosity.

That said, it is always informational to see what the releaser thinks about their own model, but even as it seems obvious that they will think it is overall better, it may not serve your own needs or preferences. This is all to say that truly, the only way to know for sure is to try and compare. It seems like I could probably write an article on my specific findings with comparisons and all, but that sounds like something a more motivated individual would be keen to do ; )

@MysteryWrecked I agree with "the worsening" of the latest releases in the most downloaded photorealistic models (I'm not talking about dream real which i have still to test). The worst offender is (most of the time) lora/embedding compatibility. I think sometimes models peaked around July-June in that regard.

@ElectricDreams Yes, conditioning is a factor too. Not only lora/embedding response, but also prompt adherence (which is usually emphasized in the better models), and even controlnet application can vary widely between models. Although that probably merits it's own study, and I really didn't look at it specifically with this round of tests.

I also forgot to emphasize the "different but equal" factor, where a few versions output nearly the same, and I dropped the ones with less detail quality. Overall, that left me with more unique models that are still at the highest quality, which is what I set out for in the first place.

@MysteryWrecked I agree too. Humans v1 was amazing at the time, V2 was not better. And similarly, mixing in the truly awful subReddit model at a low 10% strength drastically improves realism in even the latest models, getting rid of that hyper shiny/HDR crap.

Recently discovered the truality engine is also ok. Would be curious to know what your list is!

I do fine the epic negative embedding can have a huge improvement on all models, although it needs diluting with five+ ordinary words, cartoon, sketch, drawing, low quality, etc to not give boring samey composition.

@AndromedaPerson I use both negative epic embeddings plus epicphoto as a positive, and a couple epic loras together. The results are striking, especially with the epic models themselves (including subsequent mixes). I did stop using the epicrealism positive because it seems to change the output composition unfavorably. Anything that says Epic in the name is in my top (photo) models list, as well as most anything with Cyber in the name. Some interesting standouts also include the Acorn models, Colossus, Devilish, DuchaitenJourney, Hexrun, ICBNP, Juggernaut, Pirsus, Shidoshi, some of the recent Stableyogi models, a model called V77, and Vendorealistic to name a few. I actually run a batch of 80 models total, including multiple versions, and in batches of 3 I get about 1k gens for a concept if I run it 4 different ways. At resolutions up to 1152, this takes a while on my old machine, but the gens are consistent while being diverse, and many truly indistinguishable from a photo. I've been a little too obsessed with SD, kinda like the WoW kids back in the day lol, but I like to say I never do anything halfway :P

@MysteryWrecked thanks for the info! With any embedding if you use the form [:embedding:2] or 3 it means it doesn't kick in till the 2nd/3rd third step, so will affect composition even less but still give a boost in texture realism :-)

@MysteryWrecked Have you tried Analog Madness v7.0 I think is the strongest checkpoint so far.

@ElectricDreams Yes, I have 6 and 7, the newer versions are definitely in the top quality range, I would have to do a specific comparison to rank them. At the top end, differentiation is key. When you can run a batch of models and recognize one just from the thumbnails, you know it's special. I haven't had AM stand out in any particular way, but there's definitely nothing wrong with it.

By the way, for anyone interested, I did publish an article with some ranks for photo models. I actually ranked this #1 overall after some evaluating. The full list is here: Top 100 Photo Realistic Models (Maybe) | Civitai