I wanted to create a new model, starting from scratch to try not to be influenced by my previous models 100% focused on imitating reality within the possibilities and virtues of pony, I hope it comes to complement the models that I shared previously and that you try to prove it
The VAE is not baked, use SDXL_vae
Clip skip: 1
Upscaler: Im using Remacri_original, upscale 1.5 Steps 10 denois 0.35
Width and height: 896*1152
CFG Scale: 5
Steps: 35
The sampling method i use is DPM++2M Karras but try with others if you want
Description
FAQ
Comments (53)
Looks too good to be true, i didnt even download, but your samples:)
Here is the real-world case sample. Text-to-Image + HiresFix + ADetailer (yoloworld + face).
You can evaluate automatic detailing. The only thing to add is to use manual inpainting.
https://civitai.com/posts/10253122?returnUrl=%2Fmodels%2F1031397
Which is better, score on or off?
its better to use score tags, i have another model for No score
A good start for photorrealism 👌🏻, the only problems I've found so far is some deformities in the eyes and repetitive female faces
For the eyes you can try with adddetailer
@Sateluco It clearly has eye problem. Thats strange. I dont remember, your previous ones had this problem.
On every sampler, low or high CFG, even with 40 steps eyes are not good.
I wont use adetailer, its shit. Though hires can fix some of the issue.
@damarienezel296 Ty for your feedback, im gona work to fix the problem
Amazing. Awesome for portraits. Once you start to zoom out things are ok too. I generate in Forge on bnb-nf4, DPM++ 2M Karras CFG 5-7 and then Hires with 4x_NMKD-Superscale-SP_178000_G, Euler A, Uniform at 2x with 0.05 Denoise to fix the eyes for anything that was not a portrait. That way it does not smooth out details too much.
or you can just use Adetailer
@iheartlumina Every serious checkpoint with right setting produce good eyes out of stock nowadays, hires fix is not excuse.
@suveermarkon910 what is this serious checkpoint and where can I find it lol
@iheartlumina every realistic
@suveermarkon910 I think sarcasm is generally lost on Civit users. If you are subjectively fine with the results you get, continue doing what you are doing. I've met people who were ok with three hands and six fingers if their futa had a big horsecock and a rack to match. If you would rather trade in a bit of texture for better eyes, or use an extension, then follow my advice. Or maybe you have an even better solution.
@iheartlumina You were sarcastic. I thought, i was...hm
@suveermarkon910 well played, but you need a better line because half of the people around here say that seriously
@iheartlumina what i wanted to say, this checkpoint has not the classic eyeproblem, like in 1.5 era, but some deformity, and even hires cant cure it 100%
@suveermarkon910 interesting, I try majority of the realistic Pony checkpoints that don't outright look like plastic in sample pictures and can't say this one has a problem with eyes compared to others.
@iheartlumina Hm, i would say the best description for eyes is uneven and lazy. And his previous ones had not this problem.
My checkpoint folder is filled with your models dude. You are doing a great job. What is the difference between this one and satponyv5? they are both good realism models. Is there any reason to choose one over other?
I checked and compared new version with XenEro v1.7. In this version I can highlight better expressions, poses and micro-detailing like hairs and small scratches and glares on solid surfaces. Look at 4 images grid by Lizard. IOR version is quite good.
@homoludens thanks for clear info, checking Lizard images rn, it is helpful to see comparison for the same image with same seeds without downloading lots of 6.46gb files
I know sometimes we have to save space with a lot of checkpoints, with this model I try to made a fresh start, different expressions, races, better skins and many other details, I can't say it's better than SatPony original but is different
@Sateluco I feel it is better. At least, more interesting. Prompt adherence looks the same or slightly better. Fingers / hands needs to be improved (not ideal).
@Sateluco I am keeping them both plus satpony journey. Journey is my fav one but sometimes I have trouble with faces. The style is x10 better but faces is inferior compared to satponyv5 or realityIOR. Do you plan to create a new version for satpony journey soon???
@espinozaa yeah, I'm always working in all my models and I want to create a fine journey version ;)
Good start, but at the moment, for me personally, Sa Pony v5 provides better image quality. Good level of realism, better eyes, smiles, image clarity. SatPony-reality certainly deserves the right to exist. But at the moment it does not reach the level of quality set in v5.
I know you are the king of SatPony V5, this is the first approach to a different form of expressions, your feedback it's very appreciated to make a better model, thanks a lot
@Sateluco More of a fan than a king )) Thank you for the work you've done)))
This is good, definitely more real, though I still find Journey gives the most epic results by far. Journey looks like Flux but with all the Pony to have fun with.
I try this models to imitate reality but for more details I love journey to
@Sateluco Gosh, after more testing I found this model to handle Lora's really well! What it may have lost from Journey natively, it makes up for it having no issue with you throwing Lora's at it and still producing banger images! Wow, most checkpoints turn to mush with this kind of Lora abuse :P
@udf2wozza please share your images, I try with a lot of loras and examples but the real scopes came from the community
@Sateluco for those who have lack of SSD space you can produce FP8 version of your model for ComfyUI. What you need is just use commandline argument to start ComfyUI in FP8 mode for UNet and Text Encoders, then use Save Model node to save your SDXL checkpoint into 3-4Gb safetensors.
Let me know if you need precise instructions - it is quite easy. I tried it in one of the XenEro versions.
Ty for the help, it gonna be very helpful including me, in order to made my models I need to use a lot of TB
1. Go to ComfyUI folder
2. $ python3.10 main.py -h
usage: main.py [-h] [--listen [IP]] [--port PORT] [--tls-keyfile TLS_KEYFILE] [--tls-certfile TLS_CERTFILE] [--enable-cors-header [ORIGIN]]
[--max-upload-size MAX_UPLOAD_SIZE] [--extra-model-paths-config PATH [PATH ...]] [--output-directory OUTPUT_DIRECTORY] [--temp-directory TEMP_DIRECTORY]
[--input-directory INPUT_DIRECTORY] [--auto-launch] [--disable-auto-launch] [--cuda-device DEVICE_ID] [--cuda-malloc | --disable-cuda-malloc]
[--force-fp32 | --force-fp16] [--bf16-unet | --fp16-unet | --fp8_e4m3fn-unet | --fp8_e5m2-unet] [--fp16-vae | --fp32-vae | --bf16-vae] [--cpu-vae]
[--fp8_e4m3fn-text-enc | --fp8_e5m2-text-enc | --fp16-text-enc | --fp32-text-enc] [--force-channels-last] [--directml [DIRECTML_DEVICE]]
[--disable-ipex-optimize] [--preview-method [none,auto,latent2rgb,taesd]] [--preview-size PREVIEW_SIZE] [--cache-classic | --cache-lru CACHE_LRU]
[--use-split-cross-attention | --use-quad-cross-attention | --use-pytorch-cross-attention] [--disable-xformers]
[--force-upcast-attention | --dont-upcast-attention] [--gpu-only | --highvram | --normalvram | --lowvram | --novram | --cpu] [--reserve-vram RESERVE_VRAM]
[--default-hashing-function {md5,sha1,sha256,sha512}] [--disable-smart-memory] [--deterministic] [--fast] [--dont-print-server] [--quick-test-for-ci]
[--windows-standalone-build] [--disable-metadata] [--disable-all-custom-nodes] [--multi-user] [--verbose [{DEBUG,INFO,WARNING,ERROR,CRITICAL}]]
[--front-end-version FRONT_END_VERSION] [--front-end-root FRONT_END_ROOT] [--user-directory USER_DIRECTORY]
3. You need options: --fp8_e4m3fn-unet --fp8_e4m3fn-text-enc
Then add the CheckpointSave node into your workflow with 100500 LoRas and Merge Model nodes ;)
https://comfyui-wiki.com/en/comfyui-nodes/advanced/model-merging/checkpoint-save
ComfyUI, when it loads the checkpoint, converts everything into the bits specified in commandline args. So, it will same Text Encoder and UNet with the bitness you specified in these args.
For example, you can save bf16 Unet with FP8 Text Encoder.
But keep in mind that A1111 don't support the hybrids like this. It don't support any FP8 out of the box.
@homoludens Forge support it? Has it any speed increase with 8gb vram?
@suveermarkon910 Have no idea about Forge. Didn't use it. FP8 should take VRAM similar to SD 1.5 but show much better result. Not sure about speed. It depends on CUDA capabilities of your GPU. For example Tesla P40 has CUDA 6.x version and it is SO SO MUCH slower than RTX 3060 with CUDA 8.x.
But in general you will be able to have bigger BATCH SIZE because of lower VRAM consumption. OR you will be able to use higher resolution without upscaling.
@homoludens Thats sounds good. Just googled, and Forge supports too with an argument.
and should work on a1111 with same argument
Both supports it for long time.
@suveermarkon910 Few weeks ago I tried to load FP8 SDXL model in A1111. And there was an error in console that A1111 cannot use FP8. Don't remember exactly what, Text Encoder or UNet, but was unable to use it in A1111.
I didn't the commandline argument for A1111 with FP8 support.
@homoludens I just googled it, but with fp8e5m2 it should work, at least on forge.
A1111 article:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14031
@suveermarkon910 Thanks for making me know. But
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/15566
Looks like this doesn't work on linux.
@homoludens Yes, and i remember, there is a 4gb Pony here, and it worked out of stock.
@homoludens Ah, ok. But masses use windows, so its not about you:) You can use comfy.
@suveermarkon910 Anyway, this new info makes me think about trying FP8 again...
@suveermarkon910 Inpainting more convenient in A1111. But I'm still thinking to build my workflow on ComfyUI with GIMP or KRITA
After a bit more testing, I do see some eye deformities, and face deformities in general if there is umm... liquid on it or certain emotions are showing. Certain angle tokens seem to increase the issue, like "dynamic angle"