Type | |
Stats | 1,277 |
Reviews | (163) |
Published | Nov 7, 2024 |
Base Model | |
Hash | AutoV2 9C15511C75 |
If you don't feel like reading the description, at least refer to the short version.
EDIT: Removed early access. Buzz is still welcome though. Now also available for online run on tensor until Civit lets me turn on online generation. (Only allows for 25 steps though, 40 is recommended). Use dpm++_2s_ancestral with the normal scheduler for more realistic images.
UPDATE: Version 3 is here! it's even better than I expected initially. Try it out, I'm sure you won't regret it.
Way better backgrounds, better creativity, better prompt adherence. I also managed to keep a lot more of the clothing knowledge of built in characters as compared to other realistic models. It's truly incredible what Pony models can be made to achieve, and it will surely not be obsolete any time soon, at least in terms of realistic checkpoints. The outputs are raw images from the model with a hires fix.
This final version achieves what my vision for a realistic Pony model was. I've improved the sampler and scheduler compatibility by a lot, but I'd still recommend using SGM_uniform/normal/simple as your scheduler. I personally use Euler A/dpm++_2s_ancestral or DPM++ 3M SDE with SGM_uniform/normal/simple. The NSFW capabilities are unmatched, but this one can give incredible SFW results if prompted for.
Short version for lower attention span (Applies strongly for v1 and v2, and somewhat for v3 too):
Use Euler A or DPM++ 3M SDE (v3 only, older versions require Euler A).
dpm++_2s_ancestral takes longer, but gives very good results too, maybe the best but it's too slow for me.
Don't use karras, AYS, gits. Only SGM uniform/normal/simple works as intended.
I've started to notice SGM_uniform smoothes things a bit. Use Normal if you need more details. But NO - AYS or Karras or Exponential.
Score tags are still useful, but not necessary. Removing them makes the model more creative and less prone to 1girl. Preserves most of the quality too.
VAE is included
If you want to support the creation of more models like this, or you just like what you see (I could really use it, very broke at the moment): ko-fi (｡♥‿♥｡).
Make sure that you read the generation parameters described further down in order to get to the quality of the showcase gens. No adetailer or other img2img techniques have been used for the gallery images made by me, except for one hires fix. Also, the model works so good by default that I haven't considered or had the chance to test out LoRAs, so I request users to do so and post their images to the gallery. I humbly request that no one spams the gallery though.
This model is a result of an experimental DARE/Block merge between different Pony models. It strives to get the perfect balance between "Ponyness" and "Realism". VAE is included.
The prime objective of this series of checkpoints is to achieve the (subjective) perfect balance between the capabilities of anime models and real ones, while still retaining their individual knowledge. I will keep updating and making this checkpoint better with the experimental merging technique that I'm working on. Documentation on DARE-Ties merging of diffusion models is nonexistent and that makes this way harder to work with.
UPDATE (v2) (also changed checkpoint name): v2 should work really well in many different situations, while still maintaining a good amount of realism. The showcase images should portray the quality of the compositions and prompts that this model can handle. in general this version can do all the stuff that most popular realistic models can do, and then some. The NSFW capabilities are also far better than I could have imagined.
Generation Parameters (VERY IMPORTANT):
Only use "EULER A" with SGM UNIFORM. Nothing works better for this one, even after extensive testing. (v2 and v1)
Use 30+ steps. Do NOT use any other scheduler apart from SGM UNIFORM. Don't use Karras or AYS or exponential. EDIT: The newer version might handle different samplers, but the scheduler still has to be SGM uniform/normal/simple.
Simple and Normal work, butSGM uniformmakes the best gens.Hi-res fix (normal or UltimateSDUpscale) is very much recommended. Half the steps of the original gen, denoise at 0.28. Euler A + sgm_uniform again. I also recommend Ultrasharp 4x or NMKD Siax as the upscale model.
CFG can be between 4 to 10. Use higher for more creativity, lower for more realism. I use AutomaticCFG's hard mode in ComfyUI, so I prefer CFG 9 for my gens. Normal CFG usage might vary.
UPDATED Merge Recipe (not exact as I am still working with random values and iterating from there):
Version 3 changes (Each pass was made with merges into different layers at small ratios until I felt like continuing) -
First pass of Quintessence V2 with Magica Pony Realism - FireMagica.
Second pass of the resulting model with Zonkey - Realism v6.
Next pass of resulting model with DucHaiten-Pony-Real - v2.0.
Resulting model merged with Pony Realism 🔮 - v2.2.
Resulting model merged again with BeMyPony - Photo - v1.
Resulting model merged with Opticon v2-rc with Zonkey v6 as the model mask.
BeMyPony - Photo - v1 merged again into different layers this time with Opticon v2-rc as the model mask.
The previous step again, but this time with the models reversed and targeting different layers and ratios.
Version 2 changes -
Merged about 15% of Illustrious x Pony Mix into different layers. The Illustrious clip has also been merged with Quintessence V1's clip using the same DARE method.
Merged Opticon v2-rc at arbitrary but significant values I unfortunately don't remember.
Version 1:
Tool used: ComfyUI-DareMerge
Base - Quintessence V1 (DARE-Ties merge between TAME Pony v2 (Base), SnowPony v1, WAI-ANI-NSFW-PONYXL v8)
The resulting final version is everything I wanted from a Pony realistic model. Enjoy!
Keep in mind that I am very new to this, and if you love this model make sure to check out the recommended ones. Special thanks to @zyxt99565 for making the TAME Pony checkpoint, which I used as the first version base for this one. Do give it a try too as that might be one of the best "raw realism" SDXL/Pony checkpoints ever released.
I wanted to augment some of the gritty realism it provides with a bit of the more dynamic Pony concepts according to my personal preferences. DARE-TIES merging is incredible and I experimented some stuff while making this. Thus this model was born.
As of now, this model is a merge that uses the following incredible models in arbitrary ratios (credits to the creators):
Also, the prompts for the showcase images have been 'borrowed' from some of the most influential models on the site. This might help you to compare the images to those models. My prompts for the showcase images are just random ones picked from other galleries, and you can surely use better ones while actually genning.
The clips have also been DARE merged (not on every part of the merge), keeping some of the prompting styles from the above models consistent. Model works with both pure Pony type prompts, or a mixture of both Pony and SDXL prompting styles. Experiment yourself. I still recommend keeping the score tags in your prompts.
Until CivitAI turns on the online generation for this model, feel free to run the model here (I have no other option but to host it here, until Civit allows the model to run onsite). There are also other intermediate versions there, that I didn't feel worthy enough quality wise to post here:
-RUN ONLINE-
Most resolutions should work well. Feel free to discuss problems and doubts in the comments.
If you want to support the creation of more models like this, or you just like what you see (I could really use it, very broke at the moment): ko-fi (｡♥‿♥｡).