Emulate NovelAI

[AIAMIAUTHOR]

Disclaimer: I have no relationship with Automatic1111/Voldy, i take all responsability for this discussion and everything shared here

compiled, edited and tested by aiamiauthor

Newsfeed https://rentry.org/sdupdates author: questianon !!YbTGdICxQOw (malt#6065)

• November 22, 2022

Anything 3.0 VAE is just the NovelAI VAE but renamed

• November 11, 2022

NAI's "Variations" feature does (by enhance anon): Alright, variations is really similar to enhance. It sends it to img2img with strength hardcoded @ 0.8, and then increments the seed by 1 for each variation given. Nothing super special.

NAI's "Enhance" feature does (by anon): It upscales the image with Lanczos (defaults to 1.5x, which is the max), and then sends it to img2img with [whatever sampler you specified] @ 50 steps, with the denoising strength ranging from 0.2 to 0.6 (this is the "Magnitude" value that NAI shows, ranging from 1 to 5). It's like a much more expensive version of SD Upscale, which does it as tiles to save VRAM, and instead this does it on the whole image at once, so it requires more VRAM.

• October 20, 2022

List of image manipulation (img2img) methods and their implementation state #2940

No Man's Guide to GPUs
https://docs.google.com/document/u/0/d/1lF9_5MIhALo7xCxKpQCZNL_jrJdUHYgJ3prET5yC1rI/mobilebasic

• October 15, 2022

https://rentry.org/stablediffgpubuy
Stable Diffusion benchmark - GPU - Spreadsheet
https://docs.google.com/spreadsheets/d/1Zlv4UFiciSgmJZncCujuXKHwc4BcxbjbSBg71-SdeNk/edit#gid=0

Don't be suckered in by the P100 appearing to have doubled rate fp16, pytorch doesn't seem to use it.
https://discuss.pytorch.org/t/cnn-fp16-slower-than-fp32-on-tesla-p100/12146

• October 14, 2022

NAI Quick Start Guide
https://rentry.org/nai-speedrun

Every graphics card model has a slightly different result.

#2017 (comment)

• October 13, 2022

[Textual Inversion] [Dreambooth] [Hypernetworks][Aesthetic Gradients]

[Hypernetworks]
https://rentry.org/hypernetwork4dumdums
[Textual Inversion]
https://rentry.org/textard
https://pastebin.com/dqHZBpyA
https://github.com/nicolai256/Stable-textual-inversion_win
[Dreambooth]
https://rentry.org/simple-db-elinas
https://github.com/huggingface/diffusers/tree/main/examples/dreambooth#training-on-a-8-gb-gpu
https://colab.research.google.com/github/ShivamShrirao/diffusers/blob/main/examples/dreambooth/DreamBooth_Stable_Diffusion.ipynb
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb

Benchmark Xformers

Xformers .bat --xformers or --force-enable-xformers
NO --precision full --medvram --lowvram --no-half-vae --opt-split-attention --opt-split-attention-v1
NO Hypernetworks
NO v2.pt

XFORMERS Build Dev 10/13/2022

GPU ( ? )
CUDA 11.8 (mid)
PYTHON 3.10.6 (mid)
cuDNN 8.6.0 (minor)
NVCC 11.8 (minor)
Driver Version: 522.25 (minor)
Pytorch-cuda=11.8 (minor)
whl/cu118 (minor)

regarding WHL: Wheel has to match your python version (3.8/3.9/3.10) and your cuda version (11.6/11.7/11.8) - if either mismatches it won't work.

regarding GPU : A theory I have had myself for a little while now. And if that is the case it would make sense that Ampere cards would get closer to what NAI produces, as I image that NAI is probably running their backend on some Ampere based cards, like a bunch of A100's or something like that.

GPU/Xformers: #2017 (comment)

Participants trying to solve EULER A - SFW - CURATED - ENSD 31337:
Elevory, RcINS (CCRcmcpe), wfjsw

1. EULER A - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2 --xformers
2. EULER A - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2 --force-enable-xformers

Benchmark Vanilla

Vanilla .bat

NO --xformers --force-enable-xformers --precision full --medvram --lowvram --no-half-vae --opt-split-attention --opt-split-attention-v1
NO Hypernetworks
NO v2.pt

NAI - EULER - NSFW - FULL

NAI - EULER A - NSFW - FULL

NAI - EULER - SFW - CURATED

NAI - EULER A - SFW - CURATED

EULER - NSFW - FULL - PRUNED - VAE - EMA FALSE - ENSD NO - CLIP SKIP 2-

EULER - NSFW - FULL - PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2-

EULER A - NSFW - FULL - PRUNED - VAE - EMA FALSE - ENSD NO - CLIP SKIP 2-

EULER A - NSFW - FULL - PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2-

EULER - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD NO - CLIP SKIP 2-

EULER - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2-

EULER A - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD NO - CLIP SKIP 2-

EULER A - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2- ⚠️

ERROR ⚠️

• (Crash) Euler x EMA TRUE
• Euler A x EMA TRUE = no crash but noise ( expected since prunned = ema true?)

EULER A - SFW - CURATED- PRUNED - VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2- Enable quantization in K samplers

(Leg looks more like original)

EULER A - SFW - CURATED- PRUNED - VAE - EMA TRUE - ENSD 31337 - CLIP SKIP 2-

(Re-pruned?)

EULER A - SFW - CURATED- PRUNED - VAE - EMA TRUE - ENSD 31337 - CLIP SKIP 2- Enable quantization in K samplers

(Re-pruned?)

EULER A - SFW - CURATED- PRUNED - NO VAE - EMA TRUE - ENSD NO - CLIP SKIP 2-

• (Re-pruned?)

EULER A - SFW - CURATED- PRUNED - NO VAE - EMA TRUE - ENSD 31337 - CLIP SKIP 2-

• (Re-pruned?)

EULER A - SFW - CURATED- PRUNED - NO VAE - EMA FALSE - ENSD NO - CLIP SKIP 2-

EULER A - SFW - CURATED- PRUNED - NO VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2-

EULER A - SFW - CURATED- PRUNED - NO VAE - EMA FALSE - ENSD NO - CLIP SKIP 2- Enable quantization in K samplers

EULER A - SFW - CURATED- PRUNED - NO VAE - EMA FALSE - ENSD 31337 - CLIP SKIP 2- Enable quantization in K samplers

• October 12, 2022

Issues: https://rentry.org/sd-issues
Troubleshooting Euler : https://imgur.com/a/DCYJCSX
Troubleshooting Euler a: https://imgur.com/a/s3llTE5

• October 11, 2022

Current State https://rentry.org/sd-tldr

• October 10, 2022

FAQ : https://rentry.org/sdg_FAQ

• October 9, 2022

AUTOMATIC 1111 WebUi- NovelAi Emulation Setup

by anons and /2017 participants

EULER - NSFW (Full) working ✔️

EULER - SFW (Curated) - working partially, tweak? what's missing? ✔️🚧

EULER A - NSFW (Full) working ✔️ eta noise seed delta option added.

EULER A - SFW (Curated) - working partially, tweak? what's missing?syntax? ✔️🚧

NAI preset seed: 31337. Sample added.

Brackets/NAI = Brackets + Backslash/Automatic1111. Sample added.
If there's any brackets you gotta put a backslash before each one to get the same output for real.

PLMS - Testing

DDIM - Testing

(New User)
Step 0: https://rentry.org/voldy

(Old User)
Step 1: Back up your stable-diffusion-webui folder and create a new folder (restart from zero) (some old pulled repos won't work, git pull won't fix it in some cases), copy or git clone it, git init, Oct 9, 2022 last commit.

clip-embed.patch for sd_hijack.py is no longer needed. DON'T edit any files

Step 2: Get animefull-final-pruned.ckpt from stableckpt folder ( Model hash: 925997e9, source: novelaileak or novelaileakpt2) move it to webui/models/Stable-diffusion. Feel free to rename it.

Step 3: Move stableckpt/animevae.pt to webui/models/Stable-diffusion and rename it to be the same name as your NovelAI checkpoint but with .vae.pt (ex: animefull-final-pruned.vae.pt)

Optional: Rename your config.yaml to be the same name as your NovelAI checkpoint (ex: animefull-final-pruned.yaml).

It's not recommended to load the yaml because it doubles ram for no change in output, but if you really want to, rename it to [model name].yaml and place it next to your model

Step 4: Move stableckpt/modules/modules/*.pt (anime.pt, anime_2.pt, etc) to webui/models/hypernetworks
Create the directory if it does not exist.

Step 5: Run webui-user.bat, git pull or auto git pull the .bat if you want.

Optional: If you want to use v2 copy or download from https://rentry.org/nai-prior-v2 save it as anything.py or v2.py in the scripts folder, v2.pt has to be in the root of the webui folder.

prompt_parser.py fix to run v2.pt is no longer needed. DON'T edit any files.

Step 6: Load your weights (ex: final-pruned.ckpt) [925997e9].

Step 7: Ignore last layers of CLIP model : 2

After update

Step 8: Test.

Prompt by Anonymous 10/08/22(Sat)12:17:07 No.89073763

Novelai prompt:

masterpiece, asuka langley sitting cross legged on a chair
Undesired Content - None: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name
// Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590

Automatic prompt:

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair 
Negative prompt : lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name
// Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590

Prompt:

Add Quality Tags = masterpiece, best quality
Novelai : masterpiece, asuka langley sitting cross legged on a chair
Automatic: masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair

Undesired Content = Negative prompt weights
low quality + bad anatomy = low quality, bad anatomy

Negative prompt examples:

low quality

low quality, bad anatomy

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name

nsfw, lowres, (bad anatomy:1.21), bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name

EULER / NSFW (Full) - Full prompt ✔️ :

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2

EULER / NSFW (Full) - Alt prompt ✔️:

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2

EULER / NSFW (Full) - Alt prompt 2 - Mix ✔️ :

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2

EULER / NSFW (Full) - Bad prompt ⚠️:

masterpiece, best quality, asuka langley sitting cross legged on a chair
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2

EULER / NSFW (Full) - Prompt + V2.pt ✔️:

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2

EULER / NSFW (Full) - Mismatch ⚠️ : before eta noise seed delta option (ENSD: 31337)

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name
Steps: 28, Sampler: Euler a, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2

EULER A / NSFW (Full) ) + ENSD: 31337 ✔️:

masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 28, Sampler: Euler a, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2, ENSD: 31337

EULER A / NSFW (Full) ) + ENSD: 31337 ✔️: Brackets/NAI = Brackets + Backslash/Automatic1111
"If there's any brackets you gotta put a backslash before each one to get the same output for real"

Minor variations reported

Prompt by Anonymous 10/10/22(Mon)13:11:48 No.6894050 edited by aiamiauthor

Nai prompt:

masterpiece portrait of smiling Asuka (evangelion), evangelion (Hideaki), caustics, textile shading, high resolution illustration, blue eyes, contempt, feminine, woman disdain, disgust, no pupils, hair over face,orange hair,  long hair, red suit, ilya kuvshinov
Steps: 28, Sampler: Euler a, CFG scale: 12, Seed: 2870305590, Size: 512x512, No quality tags, Undesired Content: Low quality + Bad Anatomy + headdress, loli, 

Automatic1111 prompt:

masterpiece portrait of smiling Asuka \(evangelion\), evangelion \(Hideaki\), caustics, textile shading, high resolution illustration, blue eyes, contempt, feminine, woman disdain, disgust, no pupils, hair over face,orange hair,  long hair, red suit, ilya kuvshinov
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, headdress, loli,
Steps: 28, Sampler: Euler a, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 925997e9, Clip skip: 2, ENSD: 31337

EULER / SFW (Curated) ✔️🚧: working partially, tweak? what's missing?

Minor variations reported

Automatic1111 prompt:

masterpiece portrait of smiling Asuka \(evangelion\), evangelion \(Hideaki\), caustics, textile shading, high resolution illustration, blue eyes, contempt, feminine, woman disdain, disgust, no pupils, hair over face,orange hair,  long hair, red suit, ilya kuvshinov
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, headdress, loli,
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 1d4a34af, Clip skip: 2

Try 1

EULER A / SFW (Curated) - Mismatch ⚠️ : before eta noise seed delta option (ENSD: 31337)

Automatic1111 prompt:

masterpiece portrait of smiling Asuka \(evangelion\), evangelion \(Hideaki\), caustics, textile shading, high resolution illustration, blue eyes, contempt, feminine, woman disdain, disgust, no pupils, hair over face,orange hair,  long hair, red suit, ilya kuvshinov
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, headdress, loli,
Steps: 28, Sampler: Euler a, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 1d4a34af, Clip skip: 2

EULER A / SFW (Curated) + ENSD: 31337) ✔️🚧: working partially, tweak? what's missing? syntax?

Automatic1111 prompt:

masterpiece portrait of smiling Asuka \(evangelion\), evangelion \(Hideaki\), caustics, textile shading, high resolution illustration, blue eyes, contempt, feminine, woman disdain, disgust, no pupils, hair over face,orange hair,  long hair, red suit, ilya kuvshinov
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, headdress, loli,
Steps: 28, Sampler: Euler a, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 1d4a34af, Clip skip: 2, ENSD: 31337

Try 1

Try 2

Try 3 : turning on Use old emphasis implementation. Can be useful to reproduce old seeds, breaks it.

Syntax, logic, color variations?

1 #2017 (comment)

2 #2017 (comment)

• October 8, 2022

Starting point:

Prompt by Anonymous 10/08/22(Sat)12:17:07 No.89073763

Can we actually implement this?

LEFT = original leak, no vae, no hypernetwork, full-pruned
MIDDLE = original leak, vae, no hypernetwork, latest, SD_Hiijack edits and Parser (v2.pt) edits
RIGHT = NovelAI

NO longer needed. DON'T edit any files
sd_model.py to run the full model
vae.pt and .yaml ?

NO longer needed. DON'T edit any files
prompt_parser.py to run v2.pt - mod code by community, OSS license.
https://pastebin.com/5cftmKDm

NO longer needed. DON'T edit any files
sd_hijack.py for the clip-embed.patch - mod code by community, OSS license.
https://pastebin.com/SXzpTfDm

[ZeroCool22]

NovelAI is just aimed for Anime?

[ZCaliber]

As a commercial product with lewdity... It kinda has to be.

[fernicar]

As a commercial product with lewdity... It kinda has to be.

Uff! the flagship product from the rising sun land is anime, to think NovelAi aim to divert income from them so soon...
I plan to teleport to an alt-universe where Manga/Anime artist from the rising sun land could profit from their own trained models instead of NovelAi to be fair.

[danger9711]

how to apply sd_hijack patch?

[MarkovInequality]

It's been merged into Voldy's repo. Settings page below filter NSFW. Set to 2

[luckyycode]

why '2' exactly?

[grexzen]

That's the layer for Clip, you can use others and the style changes, but 2 is the one that results in the same image as in OP.

[Elevory]

Hi, I figured out how to recreate the NAI result for this image.

4chan forgot to account for the "quality tags" that NAI prepends to the prompt.

I found this post which clarifies how the quality tags work:
https://www.reddit.com/r/NovelAi/comments/xxew6a/small_update_default_quality_of_life_bug_changes/

It's just "masterpiece, best quality" to the start of the prompt. So the prompt becomes "masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair" 28 steps, euler, cfg scale 12, seed = 2870305590

Still slightly off, maybe need to tweak CFG scale or something, but it's very close.

Anyway, someone should go tell 4chan. I can't be bothered to complete their godforsaken captcha.

Also, shame on NovelAI for what they're doing to Auto.

[Desm0nt]

It's all fine for nsfw model, but for SWF it probably miss something

[Desm0nt]

If it neccessary I can upload multiple different images from SFW NovelAI. None of them can be recreated with this tutorial. Sometimes they are close with minor difference like a background, sometimes the are completely different

[Desm0nt]

With non-square NSFW from bot same thing

[pcoffey42]

When I copied the UC from the image NAI, I was able to replicate it 100%.

Something I noticed, it seems the UC value changes depending on what you put, like the fact that the image had "NSFW" in it even though it was the FULL model. Other images (specifically NSFW ones) don't have that UC generated. I think that might be the factor causing the discrepancies. To fully replicate an image you need to look in the image file each time.

That might be NAI's "secret sauce". They dynamically alter the prompt to get better images based on the keywords you enter?

The difference in gamma is just from the screencap. I can post full images if needed.

[Desm0nt]

I attached images to my post. I directly copy promt, uc and seed from this images, but it not help. Maybe they use different model or any other tricks for 768p pictures

[animebblocker]

Sorry if it is a silly Q. I got this error with the edited prompt_parser.py:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_addmm)

Not an expert with torch. Any help would be much appreciated!

[univeous]

I'm not sure if this is the right approach,
but after I replace

        self.vector_proj = nn.Linear(hidden_size*2, inter_dim, bias=True)
        self.out_proj = nn.Linear(hidden_size+inter_dim, hidden_size, bias=True)

with

        self.vector_proj = nn.Linear(hidden_size*2, inter_dim, bias=True, device='cuda:0')
        self.out_proj = nn.Linear(hidden_size+inter_dim, hidden_size, bias=True, device='cuda:0')

it works fine for me.

[animebblocker]

Thanks so much! This works like a charm.

[YNM-Kun]

Would you give me a tutorial that even an incompetent like me who knows nothing about Python etc. can understand?

[AIAMIAUTHOR]

Up.

[NekliozOld]

remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens]
         tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device)
-        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
-        z = outputs.last_hidden_state
+        outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=self.return_layer is not None, return_dict=True)
+        if self.return_layer is not None:
+            z = outputs.hidden_states[self.return_layer]
+            if self.do_final_ln:
+                z = self.transformer.final_layer_norm(z)
+        else:
+            z = outputs.last_hidden_state
 
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
         batch_multipliers_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers]

I'm not sure where I should add/delete that as I don't have those "-" lines in my sd_hijack.py file

Does someone have the edited sd_hijack.py file?

[Elevory]

https://pastebin.com/aKeFjMvN

By the way, the +/- indicate which lines you need to add or remove. You can use the lines that are not prefixed to orient yourself within the file.

[NekliozOld]

That's what I did in all the other lines but couldnt find that part, thanks!

[danielmm8888]

You don't need this any more. It's built into the settings page now, at the very bottom. Revert that file, do a git stash, do a git pull and there you go.

[dewa4211]

You don't need this any more. It's built into the settings page now, at the very bottom. Revert that file, do a git stash, do a git pull and there you go.

I got V2 loaded with the py script but the "Ignore last layers of CLIP model" is available on me, and if I use the script SD_Hijack I got an error, do you have any suggestion what is wrong?

[NekliozOld]

Also, where can I find this?

Parser (v2.pt) edits

[danielmm8888]

DON'T edit any files. DO however use this script which enables a toggle in the options to enable v2.pt or not.
Just save it as anything.py in the scripts folder, v2.pt has to be in the root of the webui folder.
https://rentry.org/nai-prior-v2

[pcoffey42]

Kind of a side question, but how can we know if the vae is working? If I named my model "AnimeFullLatest.ckpt" the correct name for the VAE file is "AnimeFullLatest.vae.pt", correct? There's nothing else I need to do regarding that?

[YNM-Kun]

A message like this is output to the console

Loading VAE weights from: D:\stable-diffusion-webui-master\models\Stable-diffusion\animefull-final-pruned.vae.pt

[slashedstar]

The cli says when the vae gets loaded

[pcoffey42]

OK, I saw that, but when I remove the VAE, I get the same results on the same seed. That's what made me not sure. Is there supposed to be a marked difference?

[luckyycode]

Seems like the latest changes broke it again
Got 2 as the clip value, Euler, 28 steps, the same seed as above, no hyper networks, VAE loaded, v2.pt loaded, tried both models (full and latest), still, output results are mismatched

Any idea?

[Andrew3216767]

If you switch off the v2.pt you will get the desired result. BUT how about making it like this with v2?...

[luckyycode]

Well, it's true, it outputs the same results now, thanks. BUT it is getting broken anyways if prompt is longer. By "broken" I mean differences are totally different, even without ()/[]. And the latest tags are totally ignored, it stops following the prompt somewhere in the middle

talking about around 75 tokens

[JC-Array]

I'm running into the same issue. There might be a step missing somewhere. Has anyone with the issue been able to identify the fix?

[B34STW4RS]

Also got doubled forearms after the latest pull, with v2 enabled.

Rando anon left this message someplace in regards to the naiprior script:

"This requires .cuda() added at the end to
vap = VectorAdjustPrior.load_model('v2.pt')
Place v2.pt in root directory"

Tested, no change.

edit2- suddenly worked after I recopied the prompt in from further up on the page. Now matching v2 current expected result.
edit3- obviously its still wrong.

[Bizori]

bruh

[D-06]

what "Stable Diffusion finetune hypernetwork" actually do since it doesnt show anything even when i pasted the leaked modules folder to hypernetworks folder

[Heathen]

I see why you didn't link it. At least it gives an idea of what it does.

[ShionYomori]

@D-06 @luckyycode @pcoffey42 @Heathen @Neklioz

Sorry for the delay, here's the X/Y of all the hypernetworks.

Settings are:

masterpiece, best quality, masterpiece, 1girl, solo, outdoors, flowers, dancing
Negative prompt: nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: [SEE COLUMN], Size: 512x512, Model hash: 925997e9, Hypernet: [SEE ROW]

"None" is what you need for NAI parity so if that's all you're after, you can ignore this setting. But it does affect things and it can be fun.
"aini" has a very strong style you might not care for that you have to really prompt to get out of (I like it 🤷 ) and going into this I would have said it has the highest consistency & quality, but from these seeds and settings, I was surprised by "pony"
I think "anime_3" is the highest quality of that set but they're all somewhat inconsistent. I generally wouldn't recommend them.
I've seen the furry hypernets be more aggressive about adding animal features so the more conservative changes here probably relate to sampler, steps, and CFG. They can get pretty wild so play around with them if you're interested.

[Heathen]

Testing to see in which cases it Hypernetworks would be a benefit.

masterpiece, best quality, Twilight Sparkle dancing in under the moonlight
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, censor
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 4242, Size: 512x512, Clip skip: 2

Not impressed by the pony results, vanilla seems to be better.

masterpiece, best quality, Lifts-Her-Tail, the lusty (argonian) maid cleaning a cupboard
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, nsfw
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 4242, Size: 512x512, Clip skip: 2

The scalie at least tried a bit more to make a scalie, altho I did have to put emphasis on argonian to get more than one result.

masterpiece, best quality, Kimi no na wa
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, nsfw
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 1822091043, Size: 512x512, Clip skip: 2

Okay, despite the AI's desire to give Spike boobs, it seems the anime ones shine when you give them actual animes to work on. Totoro on the anime hypernet is looking good. Kimi no na wa looked decent on all of them... I have no idea what happened with Cowboy bebop.

masterpiece, best quality, nekomimi sitting at a cafe
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, nsfw
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 4242, Size: 512x512, Clip skip: 2

They all look good, but I think furry_kemono is the winner for nekomimi, however, the default one is quite capable. They all did a poor job with tails, with the one exception there in the middle.

masterpiece, best quality, tiger-man walking down the street

They all kind of failed in my opinion, at least some tried to give results. Perhaps we need a better prompt for these.

masterpiece, best quality, robotic wolf girl at the bus stop

This prompt is where furry_protogen should shine and indeed, it did a good job.

So there you have it. If you want some specific things, the hypernetwork might help. The 3 anime ones for well, actual movie/tv anime. The furry ones for the specific type of furry you want to make. As for aini, I thought it was perhaps related to korean anime, ai ni means love you. But doesn't seem to be the case.

masterpiece, best quality, Yobi, The Five Tailed Fox

If anything it seems to give some nice stylized results now and then.

[ShionYomori]

Great work @Heathen , thank you for going the extra step I was too lazy to do, hahah.
You're right that Furry_Protogen is really strong and stable in its element, I'll have to poke around with that.

[baphilia]

@ShionYomori

"None" is what you need for NAI parity so if that's all you're after, you can ignore this setting. But it does affect things and it can be fun.

So, just to be clear, the furry fine-tuned model wasn't part of the leak? none of the furry pt's seem to make anything remotely similar to the site furry model, even with all the same settings and after passing the asuka test.

Also, thank you and everyone else for all the hard work.

[ShionYomori]

I'm sure someone already tested but for visibility in this thread:

Throwing stuff at the wall for Euler_A and it's not (just) the ETA multiplier in the settings tab.

Target:

Prompt: masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
"steps": 28, "sampler": "k_euler_ancestral", "seed": 2870305590, "strength": 0.69, "noise": 0.667, "scale": 12.0, "uc": "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name"

X/Y stepping through ETA values. I know it doesn't hit the .667 or 69 exact but it's clearly off base with the red legs. And when the legs are bare at lower ETA values, they're crossed the wrong way.

[Heathen]

"strength": 0.69

[ClashSAN]

👀

[Heathen]

[meowkawaiijp]

blacken

[AIAMIAUTHOR]

If you get a green/black screen instead of generated pictures, you have a card that doesn't support half precision floating point numbers (known problem on 16xx cards):
You must use --precision full --no-half in addition to other flags, and the model will take much more space in VRAM
Edit, webui-user.bat, set COMMANDLINE_ARGS=--precision full --no-half
Make sure to disable hardware acceleration in your browser and close anything which might be occupying VRAM if you are getting out-of-memory errors, and possibly remove GFPGAN (if you have it).
If this is not the issue, please show cmd errors.

[EugeoSynthesisThirtyTwo]

Did you enabled "Filter NSFW content" ?

[cmp-nct]

3090 here, I am getting consistent black screens now on img2img
It started 2 days ago with occasional black results or black frames displayed, now it's permanently.
I guess I need a restart of the PC, something is wonky

[tearxinnuan]

如果您得到绿/黑屏幕而不是生成的图片，则您的卡不支持半精度浮点数（16xx 卡上的已知问题）： 您必须使用 --precision full --no-half 除了其他标志，模型将在 VRAM 中占用更多空间 编辑 webui-user.bat，设置 COMMANDLINE_ARGS=--precision full --no-half 确保在浏览器中禁用硬件加速并关闭任何可能占用 VRAM 的内容如果您遇到内存不足错误，并可能删除 GFPGAN（如果有的话）。 如果这不是问题，请显示 cmd 错误。

Does having a GFPGAN cause a black screen even though I'm not using it?

[Seelukebair]

I was able to fix the black renders by simply using --no-half-vae. It still occasionally shows a black preview but the finished render (or interrupted render) is complete. My launch string is:
set COMMANDLINE_ARGS=--autolaunch --medvram --no-half-vae
For reference I can generate 1024 x 1024 images on my 3090 using about 4/5ths of my vram.

I originally thought it was the hypernetwork (because as of this moment it's pretty new), but when you remove the *.vae.pt file from your Stable-diffusion model folder the problem goes away. I thought this was a good compromise since it appears to be working without loading the full precision checkpoint into vram. Maybe someone can chime in with more testing... though I'm sure it will be fixed somewhere down the line.
Hope this helps someone.

[Heathen]

For anyone there wanting to test, best sampler for this model is DDIM at 50 samples. Sometimes it goes completely wrong, but the amount of details it puts on clothes and hair is quite good.

masterpiece, best quality, handsome dark elf warrior walking through a gloomy forest, closeup
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 50, Sampler: DDIM, CFG scale: 11, Seed: 3162940517, Size: 512x512, Clip skip: 2

Euler: 50 also gives decent results with some better rim lighting and some more detail here and there. But the bad parts is that the hair gets too busy instead of the clean anime style. The image gets slightly darker as well and the flair in the front of the armor is gone. The cape is more detailed, which is the good bit of this Euler version.

masterpiece, best quality, handsome dark elf warrior walking through a gloomy forest, soft shading, closeup
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 50, Sampler: DDIM, CFG scale: 12, Seed: 2999645793, Size: 512x512, Clip skip: 2

Euler: 50 version, once again, darker with a bit more contrast and some details are missing. At least this time it didn't overdo the hair.

If you want finer details in outfits and a more anime-look, go DDIM, Euler for more contrast and sometimes busier pictures.

[dewa4211]

I got error 'Options' object has no attribute 'CLIP_ignore_last_layers', I'm guessing from SD_Hijack, anyone has solution?
I got V2.PT, "Ignore last layers of CLIP model" is not available on me in the option

[danielmm8888]

You don't need SD_Hijack any more, it's built into it. If I were you I'd git stash any changes you've done and pull the latest commits.
EDIT: Also the option is called "Stop At last layers of CLIP model" now.

[dewa4211]

You don't need SD_Hijack any more, it's built into it. If I were you I'd git stash any changes you've done and pull the latest commits. EDIT: Also the option is called "Stop At last layers of CLIP model" now.

thank for replying sir, I delete SD_Hijack now, but somewhat result of the image is the one on the left, compare to 3 image on the guide, even though I got V2 loaded, do you have any suggestion for it? btw thank you so much!

[danielmm8888]

You don't need SD_Hijack any more, it's built into it. If I were you I'd git stash any changes you've done and pull the latest commits. EDIT: Also the option is called "Stop At last layers of CLIP model" now.

thank for replying sir, I delete SD_Hijack now, but somewhat result of the image is the one on the left, compare to 3 image on the guide, even though I got V2 loaded, do you have any suggestion for it? btw thank you so much!

Make sure you are using the animefull-final-pruned model, make sure to rename animevae.pt to whatever you called your model, so if your model is called animefull-final-pruned.ckpt, rename animevae..pt to animefull-final-pruned.vae.pt and put it in the same folder with your model, and (optionally) insert the config.yaml, rename it to animefull-final-pruned.yaml and again put it in the same folder as your model.

Also, disable v2.pt in the options. Sometimes it helps, sometimes it doesn't. In this case it doesn't.

[younyokel]

What do "Low Quality + Bad Anatomy" and "Low Quality" options in NovelAI website add to the negative prompt?

[daswer123]

Low Quality + Bad Anatomy
nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

Low Quality
nsfw, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

[younyokel]

Nice. I wonder why they use nsfw negative tag though it still gives you nsfw.

[Hilurex]

I need help with this error and I don't know what to do

[gmilandri]

Has anyone been able to replicate novelai multiprompts? I can't get the same results using the AND keyword

[Chibi2706]

I got this error always i launch the "webui-user.bat"

[Hilurex]

what version of python are you running? It worked for me with the version 3.10.6. try to reinstall everything with this version if it works for you

[cooperdk]

It's missing numpy which is in all Python versions. I agree something must have happened when @Chibi2706 ran webui-user.bat the first time.

[Chibi2706]

what version of python are you running? It worked for me with the version 3.10.6. try to reinstall everything with this version if it works for you

I using the 3.10 version

[Chibi2706]

Nevermind i loking and i got 3.10.6

[illtellyoulater]

... am I the only one who finds it a bit depressing that this repo's most upvoted discussion is about being able to replicate anime despite SD capabilities being incredibly more powerful than that?

[cooperdk]

Nah, I primarily play around doing naughty women who are just a bit past realistic in many ways.
I don't really care for anime/hentai. I wonder why 3 of 4 models are focused on it, frankly.

[illtellyoulater]

That's not exactly what I meant when I said "SD capabilities being incredibly more powerful than replicating anime"... but ok.

[cooperdk]

If its not about the simpler graphics and therefore simpler calculations, I admit, I have no idea what you mean.

[chawok]

As far as training what do I do with VAE?
Does it have to be loaded individually?
Can you merge it with a checkpoint instead?

[alexcantfunction]

I need some help here, when i try to generate an image i get this really long error

[Hilurex]

casuality activate -lowmemory in launcher? reinstall libraries if it does not work

[Revival8697]

It seems that your GPU ran out of VRAM. Read this.

[Dryderx]

Everything works fine BUT I always get the Asuka from the clip skip 1 (Shown in Asuka troubleshooter). No matter what i set the clip skip to, the image doesnt change, so i can't even get the clip skip 3 cursed image either lmao. Any idea why clip skip isnt changing anything?

[Simon1V]

Did you click "Apply Settings" at the top of the page in the settings tab? It wont consider the changes until you do.

[4meame]

how to fix the eyes ?

[AlUlkesh]

Hard to just guess. Post all the parameters you used, or post an image that contain the png-info, then we can look into it.

[4meame]

[4meame]

it happens after I pulled the latest commit..

[AldoPanecaldo]

Hi! I followed steps on the guide here: https://rentry.org/voldy#-novelai-setup-

Then tried to replicate the "Hello world" image but I always get the following output image:

This is the exact prompt:
masterpiece, best quality, masterpiece, asuka langley sitting cross legged on a chair
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 2870305590, Size: 512x512, Model hash: 89d59c3dde, Model: nai, Clip skip: 2

The only different things in the prompt are the model name (I renamed the ckpt file based on the name on the guide at the provided link) and the hash of the model.

What am I doing wrong?
Note that I am also using the --medvram argument at the bat file for startup, and that I have added to the Stable Diffusion folder for models/checkpoints both the "Official Stable Diffusion 1.4" checkpoint file and the "NovelAI" checkpoint model too (renamed to "nai").

[AldoPanecaldo]

Indeed, same results, but significantly slower though (4 mins), even though I might have changed something on the settings that made the generation slower

[racinmat]

The "Ignore last layers of CLIP model" got renamed to "Clip Skip", check if you have this settings correct.

[Skymaster3]

I have same problem. I don't think the problem is that there is not enough video memory. It seems to me that the problem is that the model hash number is different. Were you able to solve the problem?

[HyperLemons]

I have the same exact issue. Anyone else ever solve this?

[cooperdk]

Depending on the card you have and available VRAM it may not be possible to get the right Asuka without an upgrade.
First check if you use any setting to enable lower the cards to render, like lowvram and opt-split-attention, or xformers as all are known to have influence on the render.
In general you need at least 8 GB vram to make the right render. 4 for the model and 4 for working. A 6 GB GTX 1660 Ti is not enough, I tried.
You need a card in the nVidia 30 series to be sure to get the right result.

[cooperdk]

This can be caused by compatibility settings, for example using a Mac, a GPU with low memory (and settings to make it work) or maybe the new ARC GPUs. SD is supposed to run on an nVidia GPU work sufficient memory and nothing else. Someone made torch compatible with lots of stuff but it didn't come with any warranty. ⁣Hent BlueMail til Android ​ Den 27. jan. 2023 22.46, fra 22.46, AIA ***@***.***> skrev:

…

can you share your commit hash and your set COMMANDLINE_ARGS=? -- Reply to this email directly or view it on GitHub: #2017 (reply in thread) You are receiving this because you commented. Message ID: ***@***.***>

[AldoPanecaldo]

I am on windows and using a NVIDIA GPU, if you were referring to my problem

[cooperdk]

So I take it that you fall under one of the categories I mention: Low vram.
This happens if your GPU has less than 8 GB.
medvram and lowvram and other settings to make SD run on low memory will change the result of all models, not just nai. It circumvents some calculations in ways I don't really get but the result is slightly different than the expected outcome. You only see it if you recreate settings skrive else has shared.
This is a known issue and impossible to correct except if you buy a GPU with enough vram to actually run Stable Diffusion as intended.
I had a Geforce 1660 Ti OC and had the same issue until I got a 3060 OC 12GB.

6 GB is far from enough but can produce good results except for reproduction.

[cooperdk]

BTW you get all the reasons and explanations earlier in this thread.

[AldoPanecaldo]

I believe, at this point, the low GPU memory is the problem, as nothing else seems off...

I even tried with --lowvram but the result is pretty much the same.

It still works and produces kinda ok results for other prompts (even though sometimes inaccurate, with bad anatomy, bad hands, bad eyes and faces, and other stuff like that, even when pumping up the sampling steps) but I guess if I want to try out the full experience, the only way to make it work.

Thanks a lot for your help anyway :)

[blubblub-web]

What's exactly going on here?
I already used this model months ago and recently I reinstalled Win.
Other models works fine tho :l

[AlUlkesh]

How does that yaml file look like for you? It should start like this:

model:
  base_learning_rate: 1.0e-04

[blubblub-web]

Don't ask me why, but I wasn't able to see what's inside. (in the text editor there wasn't any character visible even if I was able to select it)
Anyways I deleted it and tried again and now it works fine :)

[raeyvolution]

Hello, thank you very much for the awesome guide!
I want to ask something. I was using NAI when it was released and I got some popular characters results using prompts easily (for example, if I enter h0shim4chi su1se1 (with proper letters of course) NAI would automatically give me a sharp-eyed blue-haired girl)
I managed to try this guide but I could not prompt said girl for example, is there a difference in models or something perhaps...? Or was there any settings/setups that I missed...? I did manage Asuka though...

[AlUlkesh]

Did you perhaps also use an embedding? Like one of these: https://huggingface.co/soknife/embeddings

[raeyvolution]

Hello, thank you very much! I just knew of the existence of Embeddings... (I just started SD yesterday...)
Ah, so we have to use embeddings like these, roger that! At first I thought I did something wrong because the NAI-styled webui somehow doesnt know these girls even though NAI could in the past... Turns out I need these. Thank you again!!

[Drift91]

Just putting this here for others that have this problem. I recall reading a while back in some guide to use --no-half without --precision full to reduce VRAM usage on my 6GB GTX 1660ti. This kind of resulted in me giving up on the AI image rendering because the results were always full of bizarre errors and deformities.

So tonight I decided to get back into it because a friend of mine was talking about his anime renders he was making with NovelAI. Again, everything was slightly messed up and badly deformed. So I decided to use the Hello Asuka test, and got this:

Turns out the simple fix was to add --precision full back to my command line arguments, which shouldn't have been removed in the first place. The result:

These are my full command line arguments, that seem to work fine on my 1660ti, as long as I don't have a game or a crapload of browser tabs open eating up all my VRAM: --precision full --no-half --medvram
If you're running out of VRAM on a 1660ti with --medvram then you likely need to close other programs that are eating up VRAM.

[catboxanon]

While this is bumped, might as well share this: https://blog.novelai.net/introducing-nai-smea-higher-image-generation-resolutions-9b0034ffdc4b

[Ehplodor]

Interesting two new samplers (sinusoidal multi-pass sampling... whatever that means) purposedly better at high resolution, at the cost of lower compute efficiency... The end of highres fix ?
Maybe implementable via https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#alternate-noise-schedules from @dfaker

[Ehplodor]

You were fast ! feature request crowsonkb/k-diffusion#55

[catboxanon]

I don't think the noise schedulers are related. The NovelAI docs mentions it runs multiple passes through the UNet, hence 'multipass' in the sampler name.

[Ehplodor]

Do you mean the "schedule" in "sine-based schedule that interpolates between multiple passes of the regular diffusion model as sampling progresses" relates to a 'multi-pass schedule' and not to the sampler's noise scheduler ?

[catboxanon]

I meant moreso the UNet multipasses are probably more important than the scheduler used. But yes, the scheduler is also related to how it interpolates between the UNet passes I believe.

[dennis-gonzales]

My Asuka's is different x))

[AlUlkesh]

@taweven107

It seems(?) to be using my GPU, but how can I confirm this?

One indicator would be the speed. Does it take seconds or minutes?

Also, for some reason my model hash for animefull-final-pruned.ckpt is [89d59c3dde] instead of [925997e9]...

The hash algorithm has changed by now, you can see that the older used less digits.

It would be helpful if you could upload your image with all information intact. The one in your post has no pnginfo. Probably a github thing. Did you copy & paste? If so try, the "attach files" thing at the bottom of the box.

[onlybrad]

I get this image too

GPU is 5700xt

arguments: --medvram --precision full --no-half --no-half-vae --disable-nan-check --opt-sub-quad-attention --opt-split-attention-v1

[johnycraft]

I am also getting this exact image how does one fix this?

[cooperdk]

I am pretty sure most of those options only work on nVidia GPUs.
Remember that this software is based on very specific nVidia tools and not guaranteed to work on neither AMD not Apple.

That said, this error typically happens when the settings and the GPU do not correspond to one another.

[johnycraft]

That would make sense since i am using a amd radeon rx series card.

but just incase is anyone with a nvidia card having this issue?