Anima Version is Available!!
Nova Anime XL
Nova Anime XL is Nova Anime: Anime/2.5D/3D checkpoint model for SDXL
Pony version aims to get the similar look as both Nova Anime and Nova Domain
All images on example were created with diffusers with custom png tags
PixAI - IL v4.0 Happy Valentine
PixAI - IL v3.0 Happy New Year
PixAI - IL v2.5 Merry Christmas
Rules
You cannot use the generated images for commercial use if it's not edited (or just turning it to black and white)
You can share images without any restriction if you don't monetize it
Advertising this model to outside is always welcome
Recommend Settings
Sampler: Euler a
Steps: 20~30
Clip Skip: 1-2
Denoising Strength: 0.65 - 0.8
(Pony)
CFG Scale: 5~7
Prompt: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, BREAK
Negative Prompts: score_4, score_5, 3d, jpeg artifacts, username, watermark, signature, normal quality, worst quality, large head, low quality, text, error, missing fingers, extra digits, fewer digits, bad eye
(Illustrious)
CFG Scale: 4~6
Prompt: masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery, {Prompt}, BREAK, depth of field, volumetric lighting
Negative Prompts: modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, signature, watermark, username, signature, simple background, conjoined,bad ai-generated
Description
Adjust some character knowledges
NoobAI EPS v1.1 + Illustrious v2.0-stable DARE applied
FAQ
Comments (40)
Hello. Do you have any plans to create a checkpoint based on Lumina, Chroma, or Qwen?
Currently no because of shortage with the sources and spaces I have
hey man, how do i get characters like cyrene if they are not in the webui tagcomplete extension? o:
Sorry if that's the case, you have to put in the prompt directly
You can see all of them on danbooru
@Crody ah so just copy paste it from danbooru? got it c:
Seems like v13 doesn't make futa well, results are totally different from prompt
Loving v13!
great model!
regarding v13, i'm having issues where the character is too small in the frame, like too much of the image is the background and it is not zoomed in on the singular character enough. anyone have prompting tips for this issue? i have been basing my prompts on the sample images on this page, but i need a bit more specific help. thanks in advance!
What's been working for me is 'close-up shot' as a prompt, or using prompts like 'face focus', 'body focus' etc.
Crody 様
いつも素晴らしいモデルを公開していただきありがとうございます
あなたのモデルをベースにしたLora学習について質問があります
sd-scriptsであなたのモデルをベースにLora学習を実行すると、以下のような問題が発生します
・真っ黒な画像を生成するLoraになる
・エラーで学習を開始できない
もしご存じでしたら、解決方法かLora学習可能な別のプログラムを教えていただく事は可能でしょうか？
ご存じなければ返信スルーしていただいてもかまいません。
■novaCartoonXL_v60.safetensorsの場合
https://civitai.com/models/1570391?modelVersionId=2329740
エラーは出ず学習は進行するが、真っ黒な画像しか生成できないLoraになってしまう
■novaAnimeXL_ilV130.safetensorsの場合
https://civitai.com/models/376130/nova-anime-xl
以下のエラーが発生し、学習が始まらない
sd-scriptsが対応していないCLIPTextModelを含んでいるためエラーが発生するようです
RuntimeError: Error(s) in loading state_dict for CLIPTextModel: Unexpected key(s) in state_dict: "text_model.embeddings.token_embedding.wrapped.weight".
■sd-scriptsのバージョン
commit 206adb643848ff27894f1e72b6987fa66db99378 (HEAD -> main, origin/main, origin/HEAD)
おそらくCLIPに対しての過計算・キー値の増加やUNetの増大が原因です
またDiffusersなどを扱ったソフトではVAEの仕様が変更された結果現在ベークされているもので黒い画像ができるエラーが確認されています
この先のモデルではこれらに対する対策をしていきます
それまでは旧バージョンを用いるのをオススメします
NovaAnime >> v11以前
NovaCartoon >> v5以前
@Crody
早速のご回答ありがとうございます！
マージに利用されたツールとの互換性かと思っておりましたが、
そういった原因だったんですね
今後のモデルではご対応頂けるとのこと感謝いたします！
また対応モデルのご案内も感謝いたします！
@fukafukagame219 以下にTeam-Cが編集したsd-scriptsを掲載します
問題となっている部分を修正して訓練ができるようにしたそうです
https://github.com/Faildes/sd-scripts/tree/support-extend
もし動作に問題がありましたらコメントをいただけると嬉しいです
I don't understand what's being said here, but thanks for posting the "Unexpected key(s) in state_dict" error. Google led me here and now I know why my LoRA training workflow was failing.
for some reason I can't load this model to train loras, the clip TE does not match the expected one for sdxl models
Here's the edited version of sd-scripts that supposed to be the fix for the issue
https://github.com/Faildes/sd-scripts/tree/support-extend
Please let me know the issue when you encountered one
Oh thanks for this, I had no idea what had broken. Just had to wipe the pc and downloaded the newest one. Just using v9 again and it's running, I don't think I'd have worked it out :)
hello! i love this model and would love to use it locally, but for whatever reason when i download it and hook it up to Easy Diffusion, i always get the error
Error: Could not load the stable-diffusion model! Reason: 'Embedding' object has no attribute 'wrapped'
and i can't find an answer for it. if anyone knows how to fix it, i would appreciate some help
I am having the same problem, and have not found a fix either.
I am commenting to hopefully attract attention to this issue.
v13 look more orange than v12
v13 is amazing, great update!!
Very useful update
Love this one, fantastic work!
When using the novaAnimeXL_ilV130.safetensors model for LoRA training, the following traceback occurred:
Files from https://github.com/Faildes/sd-scripts/tree/support-extend not helpful..
Traceback (most recent call last):
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\sdxl_train_network.py", line 184, in <module>
trainer.train(args)
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\train_network.py", line 234, in train
model_version, text_encoder, vae, unet = self.load_target_model(args, weight_dtype, accelerator)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\sdxl_train_network.py", line 43, in load_target_model
) = sdxl_train_util.load_target_model(args, accelerator, sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, weight_dtype)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\library\sdxl_train_util.py", line 43, in load_target_model
) = loadtarget_model(
^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\library\sdxl_train_util.py", line 82, in loadtarget_model
) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype, disable_mmap)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\library\sdxl_model_util.py", line 277, in load_models_from_sdxl_checkpoint
info1 = loadstate_dict_on_device(text_model1, te1_sd, device=map_location) # remain fp32
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\lora-scripts-v1.12.0\scripts\stable\library\sdxl_model_util.py", line 166, in loadstate_dict_on_device
raise RuntimeError("Error(s) in loading state_dict for {}:\n\t{}".format(model.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for CLIPTextModel:
Unexpected key(s) in state_dict: "text_model.embeddings.token_embedding.wrapped.weight".
Did you solve it? I have the same problem in another Nova model =/
How do you control hair style with this model ? I seem to get the same whether I add short or long hair to my prompt, or using a specific hair style.
Regarding the comments, the next version (v14.0) won't have the embedding key issue
This version will come out the first day of the December
Oh, the next one is coming out already? The V13 is responsive and is the best model.
@kuronekoAI It has the problem with the training
I resolved the issue on the next version
I'll fix the VAE issue in latter versions
Fixed VAE issue as well
I have confirmed that the errors during LoRA training no longer occur in version 14.0.
Thank you for your support.
What upscaler do you use to get such clean results? Mine look more blurry, especially the backgrounds.
Nearest exact, x1.5, 16 steps and 4.5 CFG
4x-AnimeSharp upscaler model can sharpen images
Version 14 is looking very promising so far. Plays with style Loras very well, the tones look very good.
Not wanting to be a ass but being already... Its way past time to have that dataset updated huh…..
The color tone of v14 is weak compared to previous versions.
To be honest, I think you should update most of the old datasets. Maybe you could also try integrating the chenkin noob model that comes with the latest dataset?
I already did with the v14.0
I feel like v14 is too clean. I prefer the deeper v13.