Dream with your Eyes Wide Open
Please enjoy PhotoSomnia, a general-purpose model merge blended and adjusted with an objective of emulating film photography, with some success.
Q. Isn't this just another model blend?
A. Short answer - No, this is the best photorealism model you'll find today.
Long answer - Yes, imo the people who are training/fine tuning and are sharing their models/embeddings/etc. are the real heroes of the AI art community! But this is an excellent photorealism blend for a few reasons - it's been the most responsive to prompts than any others I have tried, and seems biased towards producing photorealistic outputs. Also despite the incorporation of many nsfw models, it doesn't seem as biased towards producing NSFW outputs as many other photorealism blends, I consider this a positive.
Q. Isn't this just a rip-off of Analog Diffusion/Dreamlike Photoreal/any other model?
A. Though both of those models and many others have been involved in the blending process, it seems to produce relatively unique outputs thanks to the weird and chaotic blending workflow I've been using. The influence of certain models on Gravy Mix seemed quite clear, but this blend shouldn't be considered a direct derivative of any model other than Stable Diffusion 1.5.
Q. What models did you use?
A. Anything I could find that appears to be trained on actual photos of actual people, plus some artistic ones and top-quality photorealism merges. It's a big list appended to the end of this blurb, and I hate to admit it but I may have missed some, and also may not have used some on the list, but they're in my models folder and are quite likely to have been used.
Q. What ratios are they blended at?
Calculating the ratio of any contributing model would be impossible. My general approach since Gravy Mix has been to grab the latest photo-trained/fine-tuned model I can find and use the either the weighted sum/add difference merge at a low ratio with the Work in Progress model, reduce all the weightings across the board once things start to get a bit overdone, and continue the process with various models to steer things towards my idea of photorealism as best possible until I'm satisfied with the results I'm getting.
It's not as scientific as A+B ratio blends nor as controlled as block merging which I never found the time to experiment with. It's an open-ended exercise but I think the results are aesthetically pleasing enough at this stage for this model to be shared. I presume that AI-assisted blending tools running on more capable systems than my old mid-tier gaming laptop will soon make this sort of approach obsolete, especially with SDXL as a base.
Q. The outputs are blurry/weird/fake looking/suck :(
A. I'm ok with the softness and other quirks of the outputs. I've blended this model for myself and will share it with the community as it is, if they aren't useful for the style you are aiming for feel free to try using it for your own model blends to suit your own styles, or use this model for composition and img2img or controlnet with your preferred model to achieve the results you're after.
Q. What settings/prompts should I use?
A. Check the metadata of the demo pics for some examples.
Q. What licensing applies to this model?
A. It's my gift back to the community that has made so much possible for me. My understanding is that despite using other models to make PhotoSomnia, this is not a direct or provable derivative of any model other than Stable Diffusion 1.5, thus Open RAIL-M should apply. Whether or not the or not the publishers of the models listed below agree with the above licensing terms is between the user and them.
If your are using this model I would greatly appreciate attribution considering how much time and effort I've put into this, but it's not a requirement.
Q. Will there be updates?
A. The final version is final; While this is one of the best photorealism models available I'm no good at hyping it so it's slipped under the radar for most users.
Model Merge List (updated for latest version):
iirc these are trained/finetuned models that have been used at some stage of developing this blend:
Absolute Reality
Analog Diffusion
Bigger Girls Photorealism
Brits out for the lads
Cafe-Instagram
Doctor Diffusion
Dreamlike Photoreal v2
Emotion Puppeteer
Elysium_V1
El Zipang
Epic Realism
f222
Humans
Metagod Selfie Mix
NeverEndingDream
NextPhoto
Noise Offset for True Darkness in SD
Parailansfw
Project Photo Beta 2.0 LITE
QGO-10b
RPG v4
sldrRealism
StableGram
Subreddit V3
td-UltraReal
This Babe Doesn't Exist
This Is Real
Vodka
WoopWoop Photo 2.0Alpha
WavyFusion
Z_phyr Mix
Hassan's Blend (not actually used directly but so many of the other blends use it that it needs a mention)
...and of course the often forgotten but always fundamental Stable Diffusion 1.5.
afaik these are merged models only, all of which are wonderful for photorealism and/or other styles just as they are:
Analog Madness
Ares Mix
Art & Eros (aEros)
Deliberate v2
HARDblend
Hunter69 v1
I Can't Believe it's Not Photography
kkwIDEAL
MegaMerge v1.1
OpenDream 2.0
Perfection
PhotoMaxUltra v1
Protogen x34
Realistic Vision v1.4
Reliberate
s1dlxbrew
StablyDiffused's Magnum Opus
Sunshine Mix
TheAlly's Mix II: Churned
uber-realistic-porn-merge
Description
Photosomnia Omega
Last update as SD 1.5 looks to be headed for obsolescence. This is the version I've been using for a while. It's been amazing especially when combined with photorealism negative embeddings. At this stage I've lost track of the models integrated into this one, apologies to the contributing trainers/mergers.
How far things have come in a year, the future is going to be amazing! But this is where SD 1.5 model merging ends for me, thank you all who have discovered and used Gravy Mix/Photosomnia to make wonderful pictures!
FAQ
Comments (10)
So Final wasn't Final after all. Seems to be a trend with checkpoints lately.
I suspect that implementing Buzz has stirred a few of us into action! I'm not planning to do anything with it, but it's worth updating the model at this stage as it's a big leap in quality from "final" imo, and further development may not occur.
@grvy underpromise, overdeliver. My religion man! Love it, Just found your checkpoint, excited to experiments!
I'm finding small blue patches in some image generations - this used to be a problem with RealisticVision
You're right, I haven't seen it for a while but I do remember this occurring when I first played around with this model. I think it's more of an issue with some samplers than others, and hires fix/upscaling can help, I'll try and do some testing to make recommendations of what to do and what to avoid. Unfortunately I'm very short on time at present and this model is effectively unsupported.
@grvy Oh no worries, it's not super common and works. It's in my rotation
I've been using this version of the model for a while and seem to have finally hit a dead end for fresh realism models to integrate into it as model trainers move on to SDXL; I'm not going to bother merging or training SDXL models until I can upgrade my hardware, and I'm not going to bother training a 1.5 model or lora at this stage. Photosomnia Omega been very useful and works well with many Loras and embeddings. It's still far from perfect as backgrounds can be horrific unintentional mc escher worlds, and there's still a tendency for characters to be asymmetrical. But it does characters better than any previous merge I've worked on, and it's good enough for realism that I've built a little momentum up with my creations, I'm not trying to trick people and everything I create is either tagged and labelled as AI generated/AI art or posted to spaces specifically for AI pics - but many still think that my pics are actual photos!
This is my go-to checkpoint when working with celebrity LoRAs. It really allows the LoRA to shine. Other checkpoints try to make the celebrities "prettier," making them fall into that uncanny valley of "cute AI face."
Photosomnia allows the likeness shine through without distorting it. Not sure how it was made, but it works like a charm for me.
Glad to hear it! It was built on top of the Gravy Mix which was a mix of trained and merged checkpoints, but while developing Photosomnia I'd only merge trained checkpoints. I'd experiment with both weighted sum and add difference merges, at different ratios between 0.1 to 0.3. I'd reduce the overall strength when some aspects seemed overpowering by using Add Difference merge with the base SD1.5 model. I never got into block-weighted merging as my hardware wasn't fast enough to test it out what difference it makes in a timely manner.
I'm not sure what other SD1.5 models with a more scientific approach are capable of because I've moved on to SDXL after a hardware upgrade. I've tried a similar method with merging SDXL models and the results aren't really an improvement on the base models. It seems unlikely that I'll release an SDXL merge.
i'm using omega version and getting very nice results. great model, thank you! 👍
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.