I am getting a difference in quality compared to A1111. Is there a way to correct this?

[StrawBerryMilkMan]

I love the fork, however, the quality for realistic models is significantly different than A1111. Is there a way to fix this? Tried different cross attention optimization methods as well and none seemed to help.
Example

Prompts: Pretty swedish girl, detailed face, best quality, high quality, skin indentation, skin pores, textured skin, analog, film grain, detailed eyes, perfect mouth, 8k, uhd, 8k uhd, closed mouth, casual clothes,

Negative prompt: easynegative, worst quality, bad quality, nsfw, naked, nude,

Size: 512x728, Seed: 1274235326, Model: ProgenUberAnalogMix, Steps: 20, Sampler: DPM++ 2S a Karras, CFG scale: 7, Model hash: de3a92d7da
Tried different cross attention optimization methods as well and none seemed to help.

2.0.0+cu118 autocast half
xformers: 0.0.17
accelerate: 0.18.0
transformers: 4.26.1
device: NVIDIA GeForce RTX 3070 (1) (compute_37) (8, 6)
cuda: 11.8
cudnn: 8700

[vladmandic]

can you upload actual image generated in a1111 and using this repo so i can take a look at their metadata?

[StrawBerryMilkMan]

Someone pointed out that this fork isn't recognizing easynegative, I tested A111 without using the easynegative prompt and it gave me a result that is similar to the vlad fork.

This is the A1111 using easynegative

This is Vlad fork using easynegative

This is A1111 without easynegative

This is Vlad fork without easynegative

Is there any way to correct this ?

[vladmandic]

sorry for a dumb question, but what is easynegative?

[StrawBerryMilkMan]

easynegative is an embedded textual inversion trained to filter out low quality images. It is meant to be used as a negative prompt. It was trained using an anime style model but it works well for all models.

[vladmandic]

that answers it - download easynegative and place it embeddings folder, it will be recognized and used.
but note that default embeddings folder here is NOT /embeddings, its /models/embeddings (and you can change it via UI -> System Paths)

so i'm guessing that answers this issue?

[StrawBerryMilkMan]

Yes that fixed the issue, you have no idea how happy this makes me. I appreciate the quick response and support.

[qpl23]

I saw the reddit thread on this and did the experiment locally. My result was an image indistinguishable from OP's 1111 image (to my eyes):

The image looks so similar I thought I should generate a similar, related image to show I didn't just copy OP's image, so here's one with one extra generation step which has morphed the background telephone pole into a tree, for example:

These were generated with OP's model / prompt / params from the reddit thread linked previously. Metadata intact at time of uploading.

(Edit to add: in case it makes any difference, my install is a couple of days out of date; git log shows last commit as "Tue Apr 18 08:06:16 2023 -0400 / disable gradio queues on demand")

[myndxero]

There are a lot of reasons things can vary. I noticed that Vlad's has some of the k sampler stuff checked by default that auto1111's does not (I think, iirc). Check clip skip and noise delta too. Some screenshots of various things that could be different that would result in a different generation even with otherwise identical settings.

[Nrgte]

I'm also getting vastly different images with vladmandic than in a1111.
I've checked all the settings in both UIs and they seem identical. I have --no-half in a1111 so I enabled that in the CUDA settings for vlads.

I've checked the settings in Stable Diffusion, CUDA Settings, Sampler Parameters and I use the same VAE, model, prompts, seed.

I ddin't find a Compatibilty tab equivalent in vlad.

Does anyone know how to debug this further?

[StrawBerryMilkMan]

Vlads uses different locations for embeddings. Make sure you place embedding in /automatic/models/embeddings and not automatic/embeddings. This is what fixed it for me.

Same goes for loras. They should be placed in \automatic\extensions\sd-webui-additional-networks\models\lora

[Nrgte]

I don't use any embeddings, but I have the loras in models\Lora instead of extensions\sd-webui-additional-networks\models\lora
But adding them there as well didn't make any difference.

[myndxero]

I don't remember which ones, but the compatibility tab boxes were merged into another section. Do you discard the next to last sigma or use a different noise delta? Vae on auto or overriding built in Vaes anything like that?

…

On Sun, Apr 23, 2023 at 12:38 PM Nrgte ***@***.***> wrote: I'm also getting vastly different images with vladmandic than in a1111. I've checked all the settings in both UIs and they seem identical. I have --no-half in a1111 so I enabled that in the CUDA settings for vlads. I've checked the settings in Stable Diffusion, CUDA Settings, Sampler Parameters and I use the same VAE, model, prompts, seed. I ddin't find a Compatibilty tab equivalent in vlad. Does anyone know how to debug this further? — Reply to this email directly, view it on GitHub <#334 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2755JZYITFMMT4K3R6VJCLXCVSKDANCNFSM6AAAAAAXGSO52E> . You are receiving this because you commented.Message ID: ***@***.***>

[Nrgte]

No I don't discard the next to last sigma. But that was also the case in a1111:

Noise delta is on 0 on both as well.

I'm unsure about the CUDA settings, since those didn't exist. And I can't find the Compatibility settings. Otherwise everything seems the same.

[myndxero]

They're mixed in, compat settings, in Vlad. Idr what they were or I could help find.

…

On Sun, Apr 23, 2023, 13:02 Nrgte ***@***.***> wrote: No I don't discard the next to last sigma. But that was also the case in a1111: [image: grafik] <https://user-images.githubusercontent.com/131605234/233856751-14b35330-9ecb-426a-94ae-52bae3bc4b7b.png> Noise delta is on 0 on both as well. [image: grafik] <https://user-images.githubusercontent.com/131605234/233856836-c313d072-d77d-4840-8f56-88d750247749.png> I'm unsure about the CUDA settings, since those didn't exist. And I can't find the Compatibility settings. Otherwise everything seems the same. — Reply to this email directly, view it on GitHub <#334 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2755JZNC56HZF2Z2F64YETXCVVCHANCNFSM6AAAAAAXGSO52E> . You are receiving this because you commented.Message ID: ***@***.***>

[vladmandic]

most cuda settings in a1111 are specified via command line params (for example --no-half is cmd flag in a1111 and its a setting in ui here)