🦴 Any yellow Buzz donated to this model will go toward making it freely available for On-Site Generation.🦴
⚠️ Karras should be avoided with this model. DDPM/DDIM is recommended. (see below for additional options). ⚠️
👍 Please Rate! - Your feedback, rating or a follow are greatly appreciated. - Thankyou! 😊
Frisky Dingo🐕 is a hybrid model that blends IL composition with XL Photorealistic outputs.
(Highly optimized for DDPM / DDIM_Uniform - Karras should be avoided except in rare cases).
This is a hybrid model with a heavily IL biased CLIP, (primarily iLustMix v7), & an XL biased UNET. It can use IL, XL & PONY LORAs to varying degrees, though you may need to adjust the weights a bit, (situationally depending). Poses and concepts tend to work well while things like artist styles generally don't translate to realism, but can still be compositionally useful, in addition to pushing the model into a semi-realism or illustration / anime style. (I'll have a section @ the bottom for additional notes on prompting, LORAs & settings which I'll try to expand on it over time).
I used upscaling & a detailer in most example images, (4x_RealWebPhoto-v4 and / or 4x_foolhardy_Remacri).
V1 Suggested settings:
VAE: sdxl_vae (baked in).
Clip skip: 2 was used during merge, (setting 1 or 2 should yield same results).
Samplers/Schedulers: DDPM / DDIM_Uniform (will yield the best results consistently) For good secondary options: DPM++ 2M SDE & DPM++ 3M SDE / Heun & SGM_Uniform, or Euler Ancestral/SGM_Uniform. For low step options: (LCM/SGM_Uniform or DPM++ 2M/AYS, DPM++2Sa/AYS).
Resolution: all standard 1MP resolutions work well in portrait and landscape, (depending on scene complexity). In addition to standard XL resolutions, I often use 1024x1360, 1120x1440, 1232x1584, and occasionally 1344x1728, in my initial generation. Some scenes will benefit from the higher resolution in both composition & detail, but you will loose some prompt adherence and see more errors in the highest resolutions.
CFG: 3.8 - 7 (I typically use 4.6, -5.2 in my initial generation). A more stylized look will tend to creep in @ higher CFGs.
Steps: 32 - 38 (I use 36 most often, but I find 38 is sometimes required to nail a pose in more complex scenes).
DMD2 / LCM: I like to set the LORA strength low (around 0.6) & the CFG high (about 1.6-1.7) with 12-16 steps. This allows neg prompts to work & I find it to be the right balance of speed and output quality. (I'll expand on multi-pass refiner / upscaling once I do a bit more testing).
Prompting:
Natural language prompting supported by Danbooru tags. Generally less is more, (this model takes things very literally & has a slight learning curve, but I promise it's worth the time). For best results try to write a clear concise prompts summery in natural language, followed by tags to fill out details. A photo centric approach is best when trying to push realism. Too many IL tags associated with anime will start to push you into semi-realism, they are however effective and can be useful in composition. With a little time getting used to the nuance, you'll find the balance point where you can get an IL biased composition with photorealistic output. Quality tags should generally be put last, (look to my sample images for examples and general formatting). I'll add more on this in the notes section over time.
positive prompts - responds well to camera related tags: photorealistic, raw photo, amiture photo, depth of field, Fujifilm XF 50mm f/2 R WR lens, 35mm film, bokeh, etc.
negative prompts - I generally recommend keeping sepia in your negatives to overcome a sepia bias in lower CFGs. (artificial, anime, illustration, unreal) up to a weight of :1.4 if needed. (always best to keep weights as low as you can to achieve the scene). (board expressionless:1.2), is useful for getting away from the default XL poker-face.
Check in from time to time for for additional info: Prompting Insights, LORA Settings, Gen Settings and anything else useful I can think of, I'll be expanding on the Notes section over time. (see bellow).
(This was a passion project that I obsessed on for 3 weeks. I hope you dig it).
🗈 Notes & Tips:
(to be expanded over time),
Notes: Frisky essentially started as a "full- Realism" branch of Fabled Infusion, they have very similar ingredients but in different ratios. While testing the latest Fabled version and discovered that the CLIP had picked up some issues over the iterative versions, so I started from scratch. I followed a very similar process to what I did with FI, but I avoided the mistake of using LORA's to stabilize sub-mixes, before a final mix. (I suspect that was the biggest issue with my process). I used a couples LORAs here, but only subtly in the final step. Additionally, I rebuilt the clip in the final step. I think this is my most cohesive merge to date. It should be quite similar to what you're used to with FI, but way more photorealism biased. It also takes most LORAs better, (at least in my testing so far).
LORAs: It's best to split CLIP & Model strength if your UI allows for this. If I'm using an IL or pony pose LORA, I tend to set the CLIP at about 0.92-1.0 & the model much lower, (I test to find the point where I can get the pose, without influencing the visual output of the checkpoint), This tends to be in the 0.60-0.80 range. If you don't have this option you can just set LORA strength a little lower until you find the best balance point, (but it's very nice to have fine control over this).
Prompting: Your prompt should be structured in order of importance / what you want to see in the foreground, followed by additional details, and finally, the medium & quality tags.
FD handles specificity well, I.E "Gibson Les Paul guitar" will yield much better results than guitar.
Be aware that the model can be hypersensitive some tags & phrases. This is true across most SDXL models, but I notice it hear. For example using terms like "hyperdetailed warm skin texture with viable pores" or "freckles" can overcook your output, especially at higher CFGs. I generally use lower weights like (freackles:0.25) to compensate.
I was using "peach fuzz" in my list of skin detailed but peaches started appearing everywhere. I've switched to "baby hairs on body" or "baby hairs illuminated by sunlight". You'll likely find examples where you need to find a creative way to reword something.
(I'll add to this section and organized it shortly).
⚠️ Karras should be avoided with this model. DDPM/DDIM is recommended. ⚠️
⭐ Resources used ⭐ Creator thanks ⭐
Checkpoints:
Uncanny valley by meden - (clip only)
Perfection Cinematic v3.1 by 6tZ
GS | Realistic & Semi-Realistic CR v2.0 by GrooteS
PornMaster-Pro Realism IL v4 by iamddtla
Lustify v2.0 v4.0, OLT (Fixed Textures), & GGWP (v7) by coyotte
NatViS: Natural Vision v1.0 & v2.5 by ndimensional
The Araminta Experiment (SDXL+Flux) Fv1 & Fv5 by aramintastudio
Acorn Is Boning XL XLv1 & XLv2 by Seeker70
Anteros XXXL v1.0 by erotes_anteros
SDXL Photorealistic Mix [NSFW] v1.0 by Adahm
iNiverse Mix XL(SFW & NSFW) GuoFeng XL v1.5 by JinnGames
ASTRAL FORGE XL v1.0 by MFM_STUDIOS
SDXL Unstable Diffusers ☛ YamerMIX NihilMania by Yamer
Juggernaut XL v XI by KandooAI
LEOSAM's HelloWorld XL v7.0 by LEOSAM
Colossus Project XL (SFW&NSFW) v12c by Afroman4peace
Photonic Fusion SDXL Finalé by StecFX
PhotoArt v6.0 by OliviaRossi
Cinema Diffuso XL by BastianAI
BATCH XL ( PHOTO REAL) by batchofcookies
LORAs:
SPO-SDXL_4k-p_10ep_LoRA_webui by rockeycoss
dark (dramatic chiaroscuro lighting) by ntc
custom style Lora
(Additionally, thanks to anyone whos prompts I've pilfered for testing).
Description
⚠️ Karras should be avoided with this model. DDPM/DDIM is highly recommended. ⚠️
(see description for additional sampler options, settings, prompting guidance & additional help).
V1.0, (Strong in: prompt adherence, detail, high resolutions, text, diversity of styles & people, highly complex scenes). Handles specific details well, I.E "Gibson Les Paul guitar" will yield much better results than "guitar". (If you're willing to adjust your prompting style to this model, it can really deliver.
Frisky Dingo is essentially a "full- Realism" branch of Fabled Infusion. I followed a similar but more refined process & used many of the same models, but this merge uses the Unet from several additional XL models & I rebuilt the CLIP in the final step. You should see a lot of compositional similarity to FI, but with a much higher photorealism bias, better overall cohesion, & better LORA handling.
FAQ
Comments (12)
I use Fabled Infusion and I love what you've done with this checkpoint. Could you explain the differences between Fabled Infusion and Frisky Dingo? I can't find any descriptions of the strengths and weaknesses of your respective checkpoints; adding this information to their descriptions would be very helpful. ❤️
Np. Thanks for the feedback. They have very similar ingredients but in different ratios. Frisky essentially started as a "full- Realism" branch of Fabled. During that time I was still testing the latest Fabled version and discovered the CLIP had picked up some issues over the iterative versions, so I started from scratch. I followed a very similar process to what I did with Fabled, but I avoided the mistake of using LORA's to stabilize sub-mixed before a final mix. (I suspect that was the biggest issue with my process). I used a couples LORAs here, but only subtly in the final step. Additionally, I rebuilt the clip in the final step. I think this is my most cohesive merge to date. It should be quite similar to what you're used to with FI, but way more photorealism biased. It also takes most LORAs better, (at least in my testing so far). I'll add an edited version of this to the notes / tips section. Thanks again, let me know if you need my to clarify or expand on anything.
@prodajie Thank you for this clarification, much appreciated. Your explanation raises a question I've had for a long time regarding CLIPs. Indeed, I often use a different CLIP than the one from the checkpoint that i use. I have several L/G CLIPs, and most of the time, one of them ll provide a better image than if using the one from the checkpoint. Is this a somewhat expected pattern regading CLIPs and checkpoints?
@antarek I'm no authority, but yes, this is a pattern I see with many merges including mine. CLIPs seam to break over time from multiple successive merges. I discovered a while back, (as you mentioned), that swapping CLIPs can restabilize a merge, so I've worked that into my process. I now treat CLIP & Unet as the separate entities they are.
@prodajie Alright, thanks for your input, much appreciated.
So when you say Karras should be avoided are you trying to say that this is optimized for a particular schedule (uniform in this case)? So safe to assume AYS is also a no-go?
I suspect AYS is a good option, but I haven't tested with it yet. I've seen good results with DPM++ 2M SDE / Heun, & decent with SGM_Uniform, I think it's likely the way karras handles noise, (linearly), that's overcooking things. I'm digging out of the worst blizzard I've seen in a decade, but I'll test some more combos ASAP, hopefully tonight. If you're already setup for it though, I'd say it's well worth a try. Great question, Thanks for your interest. I hope you enjoy the model.
After 2 nights of troubleshooting (trying to satisfy all dependencies to run the required nodes on my outdated system), I'm happy to say, AYS works pretty well @ 10-12 steps, even in quite high resolutions, (could prob benefit from CFG amplification & has similar tradeoffs to LCM, but a bit less soft focus). I'm generating at 1528x1232, CFG: 2 - 3ish, 12 steps. It seems to pair best with DPM++ 2M, (faster) or DPM++ 2S a (almost 2x the gen time, but far more refined), (some differences in composition and details), I'll update the documentation after some sleep. Thanks for the inquiry, glad I can run this now. Let me know if there's any other combos you're interested in. I'll test anything, as long as I can run the nodes.
Hands down the best photorealistic model I've used, kudos. Gotta learn how to wrestle it a bit with weights and prompt order but that's part of the fun. All in all if you know how to use Illustrious you know how to use this. Also works very well with the dmd2 lora.
If I may ask, is it just me or the model seems very reluctant to do anal? Tried with weights, focused prompts and inpainting, but it struggles hard to get it right.
By comparison Fabled Infusion seems more responsive with these and other details.
Thank you vey much for the kind words & feedback. I hadn't noticed this, but I did some testing last night & confirmed what you're describing. A couple things I found (somewhat) helpful, try adding (vaginal, vaginal penetration:1.5) to your neg, and for some reason using DPM++2m-SGE with SGM_Uniform seems to increase the chances, (even then it only worked on a few seeds for me. I'll try to address this in the next version, but adding versatility without loosing photorealism is a challenge.
Thanks again for the post.
@prodajie thank you for the great model, and thanks for taking the time to test this out; I'll experiment with your solutions and see if I can get them to work.
I'm not a big site user and I'm not sure what Buzz is for, but I've sent whatever little I had your way. Can't wait to see what's in store next for Frisky Dingo and your other models.
@Jazzbanana Cheers. TY for the tip. I use yellow buzz to bid on models, (to make them available for on-site generation). Also useful for creating LORAs. I think the other 2 colors are mostly useless though.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.