Hello @FefaAI ,

The VAE should go into \models\VAE, then you can select it via the UI using Settings -> Stable Diffusion -> SD VAE.

Hello @FefaAI ,

That is one way, yes, in the case of the pruned versions here, I have baked the VAE into them, so setting it to automatic should use the orangemix VAE in the model automatically. If the colors look faded, that would indicate that the orangemix VAE likely isn't in use and you may need to tweak the setting to specifically use the VAE, or use the rename after the model name approach.

Hello @whywhynot ,

Yes, this model leverages danbooru tags.

Hello @whywhynot ,

It will use the VAE you select, it will not use both or combine them. 

@whywhynot the pruned models have the VAE baked in, the unpruned ones do not. It will only use the baked in VAE if set to automatic or none (none just doesn't try to find a VAE matching the models name). I will look at uploading unpruned ones with VAE embedded if there is a desire for them. If you are facing issues even with the pruned ones, then manually selecting the VAE should work fine as well.

Also, do be aware that the other main VAEs, i.e. SD's MSE or WD 1.4 (or their dozens of renamed instances), tend to greatly over saturate colors by comparison to the orangemix VAE. If you prefer that, then you would need to manually use one of those, otherwise using (monochrome:1.1) and (greyscale) in the negative prompt will improve the colors as well.

@laughterthoughts did you download a pruned version? Those have the VAE baked in, if you are using the unpruned version, you need to download the VAE and select it manually. In addition, use (monochrome:1.1) and (grayscale) in your negative prompts for better color.

@laughterthoughts ah, yeah, that one requires you to manually select the VAE. The pruned versions have the same quality level and include the VAE, so you can leave it on automatic for those.

@laughterthoughts if you click the dropdown next to download latest, it is available there. Here is a direct link - https://civitai.com/api/download/models/5038?type=VAE&format=Other

@laughterthoughts place it into \models\VAE and select it in the UI via settings -> stable diffusion -> SD VAE.

Hey @kurokuro ,

Negating blush seems to work for me, might need to emphasize it more, such as (blush:1.1) or even (blush:1.2). It won't 100% remove it, likely the model is over fitted on that trait, then you could easily remove it entirely via post processing if you really want it all gone. Certain skin tones might also not have as pronounced. In addition, if using an VAE like WD 1.4 or SD MSE, those will over saturate everything, making the blush more visible.

@kurokuro I'll try digging into more when I get a chance to see if there is something we are missing that might address it. The models and mix recipes involved are different from BloodOrange, so it's behavior being different makes sense.

@kurokuro apologies about the delay, but WarriorMama777 did make some recommendations on tags to try negative prompting to remove blush. These were for AOM3, but I'd imagine they should work here as well.

(blush, embarrassed, nose blush, light blush, full-face blush:1.4),

Hello @taiyakiss ,

Saturation
First, make sure you are using the correct VAE, either the orangemix one manually, or by using one of the pruned models which has it baked in (so setting SD VAE to automatic will use it).

Next, I'd recommend negative prompting (monochrome:1.1), (greyscale) to improve colors.

Finally, if that still isn't enough, I recommend using the SD MSE OR WD 1.4 VAE, which will (in my opinion overly do so) saturate the colors.

SD MSE - https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors

WD 1.4 - https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt

Style
Positive prompting realistic can greatly enhance the realism and 3d appearance of images. photorealistic and 3d are also tags to try. The amount of empahsis depends on the prompt and settings, so experiment with higher or lower amounts of it e.x. (realistic:0.8) or (realistic:1.2).

Hope this helps!

@NiConfucius

These instructions assume you are using Automatic1111, I am unable to guide for other UIs.
1. Download https://civitai.com/api/download/models/5038?type=VAE&format=Other
2. Place into ..\stable-diffusion-webui\models\vae
3. Select in the UI via Settings -> Stable Diffusion -> SD VAE

Hello @Drai ,

So there might be a couple of factors at play here. Assuming you are using the baked in VAE or the orangemix VAE seperately (as fine details can be affected by the VAE). This is also assuming you use are using the Automatic1111 UI.

1. I am using xformers (and perhaps you are too), which introduces a small amount of non-determinism, i.e. pictures can vary a small amount given the same seed, prompt, and other generation parameters.

2. I am using a setting in the UI that might slightly change results, you can find it here in the UI:
Settings -> Stable Diffusion -> "Enable quantization in K samplers for sharper and cleaner results. This may change existing seeds. Requires restart to apply."

Outside of those, what you can try is using variation seeds to introduce small amounts of change on a given primary seed. In the Automatic1111 UI, this is enabled via the Extra checkbox, then you just allow the variation seed to be random (-1) and set it to have a small amount of strength (0 being no change, 1 being completely changed, so starting at 0.05 or less might be effective).

All upscaling was done via hires fix using latent (nearest) or latent (nearest-exact). None of my examples leverage any img2img, inpainting, or post processing.

Hello @DenisKeni ,

Yes, NSFW would typically be better for SFW, as hardcore will favor more sexual images.

Quality may be better with NSFW for SFW images as well, as hardcore introduces a model which, while better for hardcore and hentai, may not be as good for cases where you do not want that.

Hello @bententhhousand ,

To replicate, you actually want to click the copy button and paste all of it into the positive prompt in Automatic1111 (replacing all of the content there if there is any), then click the little blue button with a white arrow in it that is to the right of the positive prompt input, or under the generate button.

I have also pasted the full generation parameters below, hope this helps!

absurdres, 1girl, blue hair, sports bra, fellatio, (penis:0.9)

Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (greyscale), multiple views, comic, sketch, animal ears, pointy ears,

ENSD: 31337, Size: 512x512, Seed: 174470057, Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Clip skip: 2, Model hash: cadf2c6654, Hires upscale: 2, Hires upscaler: Latent (nearest), Denoising strength: 0.6

The upscaling is likely the critical piece as to why the results are so different, but I also did not use Abyss in my positive.

Hey @Evergale ,

The best way to replicate an image, assuming you are using the automatic1111 UI, is to click "Copy Generation Data" and paste it into the positive prompt (overwriting anything currently there) in the automatic1111 UI, THEN click the blue button with an arrow to the right of the positive prompt (or under the generate button), this will automatically map all the generation parameters for you.

Hey @msiaigens ,

So, make sure you are carrying over the negative prompts using txt2img for when you do img2img as far as the quality, resolution, and color related ones. The other thing to try is the setting in the UI called "Apply color correction to img2img results to match original colors." I'm not sure the exact menu this is under, but you should be able to find it somewhere under the settings tab in the UI.

@Juan_Hernandez @dilectiogames

Make sure you are using the correct Clip Skip and Eta noise seed deltas, as sometimes those seem like they aren't being mapped properly when mapping the generation parameters. These examples expect you to set Clip Skip (aka Stop At last layers of CLIP model) to 2 and Eta noise seed delta (ENSD) to 31337, this is typical of most anime models. These settings can be found under the settings tab in the UI. At this time I am unsure of where, as I have been using them via the quick settings for so long, I have forgotten, will reply again once I find where they would be. Other than that, make sure you are using the right model, that all of the generation parameters are the same, and that no extensions are interfering and changing anything.

Hello @evelene02192 ,

The VAE would only effect fine details and color. The key thing is that you need to use booru tags in your prompts, as if you don't use the full tag, it won't understand it correctly, or at all. For cowgirl, for example, you'd want to try including cowgirl position, straddling, sex, vaginal in your positive, and negating other positions by their tag names in the negative if you see conflicting positions occurring. Great (NSFW) reference resource for them, https://danbooru.donmai.us/wiki_pages/tag_groups.

Hello @CreatorEx ,

The 2:3, e.x. 512x768, results are not being cherry picked with this model. If you are trying to directly generate at resolutions exceeding 512x512 too much, you will get incoherence, distortions, and abnormalities. The solution is to set a lower resolution for the first pass, then use hi resolution fix to upscale the image. Almost all models are trained on data 512x512 or smaller, as such they cannot effectively handle higher resolutions directly.

I can generate hundreds of images with the following settings, and not see the issues you are facing.

absurdres, 1girl
Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (depth of field, bokeh, blurry:1.1),(motion lines, motion blur:1.1), (greyscale, monochrome:1.0),
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: -1, Size: 512x768, Model hash: e892703c61, Clip skip: 2, ENSD: 31337

It is best to copy generation parameters from examples for replication (i.e. copy them from the images that have the little info icon, clicking copy generation parameters button, and pasting that into your positive prompt, overwriting anything there, using the little blue button to the right to automatically map them), simply copying the prompts and then using your own configuration will lead to unexpected results.

@CreatorEx I see what you mean, there is a better tag for framing that will capture more of their body, cowboy shot, reference (NSFW) https://danbooru.donmai.us/wiki_pages/cowboy_shot.

Try a prompt like this

absurdres, cowboy shot, (animal ears:0.1), large breasts
Negative prompt: (worst quality, low quality:1.3), (lowres), (depth of field, bokeh, blurry:1.3),(motion lines, motion blur:1.1), (greyscale, monochrome:1.0)
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2650797226, Size: 512x768, Model hash: cadf2c6654

Hello @zichen , would need more detail. Using the right VAE can help, OrangeMix / NAI for more typical anime style, and the Stable Diffusion MSE one for more realistic images. In addition, I do not use or recommend using face restoration at all, it typically mangles faces. Upscaling images with sufficient d noising can also resolve issues with eyes.

SD 2.X isn't really an upgrade for end users of models for image generation, it has some improvements, but nothing significant enough to justify it's use without some amazing new models being released. SD 1.5 ecosystem also has far more content and ongoing creation of new content than SD 2.X does.