Wanted to try it today for something, returns this error when generating. Worked fine like a week ago I think

Traceback (most recent call last):
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/call_queue.py", line 56, in f
    res = list(func(*args, **kwargs))
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/call_queue.py", line 37, in f
    res = func(*args, **kwargs)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/txt2img.py", line 56, in txt2img
    processed = process_images(p)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/processing.py", line 486, in process_images
    res = process_images_inner(p)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/extensions/sd-webui-controlnet/scripts/batch_hijack.py", line 42, in processing_process_images_hijack
    return getattr(processing, '__controlnet_original_process_images_inner')(p, *args, **kwargs)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/processing.py", line 626, in process_images_inner
    c = get_conds_with_caching(prompt_parser.get_multicond_learned_conditioning, prompts, p.steps, cached_c)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/processing.py", line 570, in get_conds_with_caching
    cache[1] = function(shared.sd_model, required_prompts, steps)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/prompt_parser.py", line 205, in get_multicond_learned_conditioning
    learned_conditioning = get_learned_conditioning(model, prompt_flat_list, steps)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/modules/prompt_parser.py", line 140, in get_learned_conditioning
    conds = model.get_learned_conditioning(texts)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/repositories/stable-diffusion-stability-ai/ldm/models/diffusion/ddpm.py", line 669, in get_learned_conditioning
    c = self.cond_stage_model(c)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1215, in _call_impl
    hook_result = hook(self, input, result)
  File "/content/drive/MyDrive/stable-diffusion-webui-colab/stable-diffusion-webui/extensions/sd-webui-cutoff/scripts/cutoff.py", line 142, in hook
    assert tensor.shape == t.shape
AssertionError

It would be nice if could give more details about the gather embeddings ? From the image from README.md, it seems that just simple concat the clip embeddings but still got an shape of 768*77, just same as the output from hide all tokens's.

I tried a different scenario, unrelated to colors.

When I apply a MASK TI to the prompt, instead of adding only the MASK, it changes the character.

ORIGINAL

<lora:Pencil-lighter:1> a lion sitting on a bench in front of a pond, a bird flying in the blue sky, (beautiful eyes:1.2), cute cartoon character, very detailed pencil drawing, (solo: 1.9), (beautiful face:1), (beautiful eyes:1)

ADDING MASK

<lora:Pencil-lighter:1> a lion sitting on a bench in front of a pond, a bird flying in the blue sky, (beautiful eyes:1.2), cute cartoon character, very detailed pencil drawing, (solo: 1.9), (beautiful face:1), (beautiful eyes:1), OVERMASKED

As you can see it completely changed the LION. Which of these words in the prompt should I add to Target Tokens, and what should the other parameters be?

repro steps

1. attempt to generate something with addon enabled and no negative prompt

error

ValueError: max() arg is an empty sequence

Just making this in case someone else runs into the same issue, just make sure to include a negative prompt and everything works fine.

when i try using the extension with a negative embedding that puts the token count over 75 it does this

works fine without it, also works fine regardless of how high the positive token count is

With A1111 v1.8.0, I get the following error on startup:

*** Error loading script: cutoff.py
Traceback (most recent call last):
File "V:\AI_programms\stable-diffusion-webui-180\modules\scripts.py", line 527, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "V:\AI_programms\stable-diffusion-webui-180\modules\script_loading.py", line 10, in load_module
module_spec.loader.exec_module(module)
File "", line 883, in exec_module
File "", line 241, in _call_with_frames_removed
File "V:\AI_programms\stable-diffusion-webui-180\extensions\sd-webui-cutoff\scripts\cutoff.py", line 13, in
from scripts.cutofflib.embedding import CLIP, CLIP_SDXL, generate_prompts, token_to_block
ImportError: cannot import name 'CLIP_SDXL' from 'scripts.cutofflib.embedding' (V:\AI_programms\stable-diffusion-webui-180\scripts\cutofflib\embedding.py)

Any ideas what this causes and if it can/will be fixed? 🤔
Also, how is the status of cutoff in regard to functioning with photo real checkpoints?

Hello, cutoff is a good extension, because the most troublesome aspect of stable diffusion is color pollution. Recently, due to work reasons, we have to transfer the workflow from auto111 to comfyUI. However, cutoff is essential. If the author or some code master has time, PLS create a cutoff node for comfyUI,

Thank you. I wish you have a nice day!

what to put in "padding token"??? hints?? please?

Great extension. also can be better.

There is no need to fill Target tokens manually. You can parse prompt, split them by "," and " " to get the first word, then fill then into Target tokens automatically.

And you can also create a color word list, as a filter, to remove non-color words, which is an option can be enabled by user.

It won't be perfect, but still be more convenient

Hello,
when using a wildcard with dynamic prompt (https://github.com/adieyal/sd-dynamic-prompts), assuming i have wild cards for colors named colors1 and colors2, the following prompt will generate random colors at generation time :

a cute girl, __colors1__shirt with __colors2__tie, __colors1__shoes, __colors1__hair, __colors2__eyes, __colors2__skirt
2

for example, it will generate this :
a cute girl, white shirt with green tie, red shoes, blue hair, yellow eyes, pink skirt

i thought it would be cool if we could enter __ colors1__ , __colors2__in the settings of the cutoff extension to make sure that any generated word for colors1 and colors2 is selected for the cutoff. What do you think ?

Anyway, thank you for your awesome and much needed extension.
Best regards,

I was testing this out in Automatic1111 SD and it looks to be pretty promising. I have a few questions though, and I hope you could add to the README after clarifying.

How does this handles multi word colours and also other non-colour adjectives? Say for example you have 1girl, white shirt, dark brown jacket, denim jeans, blonde hair. Would my target tokens be: white, dark brown, denim, blonde?

There are two things I am curious about, can it do the dark brown part or are multi-word colours not supported?

Secondly, say I'm using denim instead of say blue, or blonde/brunette instead of gold/brown, will this still work?

I also have general questions regarding some points someone mentioned on the CIVITAI promotion of this extension. Can you elaborate on the 75 token limit they might be referring to? Is that a limit on it functioning, or is it more of a performance limit and it still works despite this?

Also they mentioned the targets must have a trailing comma, e.g. white, blue, instead of white, blue, is this true? Seems like a fairly simply bug to fix and wouldn't surprise me if it has been resolved.

That's all I need to know for now, great work on this 👍 I look forward to how it will improve in the future, maybe some QoL things like auto detection/autofill like another user suggested.

I'm working with sd-webui's API + controlnet + sd-webui-cutoff to take black and white line art manga/comics, and color them with the correct colors for the characters that were detected (by gpt4o) in a given panel of a comic/manga.

This works (somewhat) well when there is a single character in the image.

But it doesn't work when there are multiple, it doesn't know which character has what color hair, color clothes, etc.

How do I solve this?

I'm looking for any solution, no matter how difficult, to try to implement. If there are multiple possible ways to attempt this, i'd like to learn about / try all of them.

Any help would be very very welcome.

Thank you

Hello, I'm having hard time figuring out if this extensions is working for me. Turning it off and on or changing the weight slider just doesn't produce differences I'm able to notice. Logs appear to be doing something, so not sure if this is intended behavior or not.

Hello, I noticed that if you only have one color, the system does not work. As soon as you put in the second color in the target tokens, it starts working. This is 100% reproduceable.

Does this work with SDXL? I tried but it fails to run with this plugin enabled.

Thanks

Hi,
How can I use it with webui api?

Cutoff targets: ["white", "black", "green", "red", "blue", "yellow", "pink", "purple", , "bronze", "blonde", "silver", "magenta"],

=>

Cutoff targets: "white, black, green, red, blue, yellow, pink, purple, , bronze, blonde, silver, magenta",

Hi
do you wish to have this extension added to the webui Extension index?
if you do, create a PR to https://github.com/AUTOMATIC1111/stable-diffusion-webui-extensions

Warning: ControlNet failed to load SGM - will use LDM instead.
[Cutoff] failed to load sgm.modules.GeneralConditioner
*** Error loading script: cutoff.py
Traceback (most recent call last):
File "/content/stable-diffusion-webui/modules/scripts.py", line 274, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "/content/stable-diffusion-webui/modules/script_loading.py", line 10, in load_module
module_spec.loader.exec_module(module)
File "", line 883, in exec_module
File "", line 241, in _call_with_frames_removed
File "/content/stable-diffusion-webui/extensions/sd-webui-cutoff/scripts/cutoff.py", line 13, in
from scripts.cutofflib.embedding import CLIP, CLIP_SDXL, generate_prompts, token_to_block
File "/content/stable-diffusion-webui/extensions/sd-webui-cutoff/scripts/cutofflib/embedding.py", line 15, in
class ClipWrapper:
File "/content/stable-diffusion-webui/extensions/sd-webui-cutoff/scripts/cutofflib/embedding.py", line 16, in ClipWrapper
def init(self, te: Union[CLIP,CLIP_SDXL]):
NameError: name 'CLIP_SDXL' is not defined

このようなエラーが発生し、項目自体が表示されず利用が出来ないようです。

Thanks for your nice work! Can I find some similar code at huggingface's diffusers? Or can you add this work to the diffusers?

Tried to disable another axtentions, change clip skip, different weight but it still doesnt work.

prompt: a cute girl, white shirt with green tie, red shoes, blue hair, yellow eyes, pink skirt
target tokens: white, green, red, blue, yellow, pink

Is it possible to train a LoRA together with an Embedding? Here are some thoughts that came to this, when training a LoRA for an object:

1. Training the entire CLIP is wrong. It is best left frozen.
2. Without learnable CLIP, we cannot change the meaning of words.
3. With or without learned CLIP, given a prompt "a photo of sks in the forest" – why would LoRA learn sks but not learn photo and forest along?
4. Generally, I do not want to learn anything except my token.
5. You could say "just use TI then!", but Embeddings are weak at learning complex concepts.
6. You could say "use regularization then!", but in this case there is no "class word" (and I don't want to introduce it); making regularization against "forest" and anything I might have in descriptions – feels wrong.
7. If it would be possible to use a learnable embedding in place of chosen token ("sks", possibly initialized with class word), then it would be more correct, because the object would be clearly stored inside this embedding and not in any other word.
8. LoRA general training should help the embedding to reach its target quicker. It's a compromise between training the entire CLIP or not at all.
9. Learning rate for the embedding should be set differently than learning rate for U-Net (or for CLIP if needed), because the best speed is yet to be discovered.

What do you think? Otherwise, I'm not quite sure how to train LoRA on something that is not a character nor a style. For example, to train a LoRA for "scar" concept: what descriptions should we choose?
Should we say "sks over eye, 1boy, …"? If so, isn't it more logical to say directly "scar over eye, 1boy, …"? But if so, how can we be sure that only the concept of "scar" would be changed, and not the concept of "1boy"?

Working on my extension Embedding Merge, I found out that if you zero-fill TI-embedding and replace a part of the prompt with it – then you will have the result completely without the affected concept, with almost no influence to other concepts in the prompt.

Actually, I implemented a pre-multiplier that you can add to any word to change it weight before the CLIP phase. So, by multiplying by zero you can get rid of anything, while still have "padding" for other words.
From my experiments, such zero-padding is working better than padding with commas, exclamation signs, with pad_token and so on, especially when merging parts with different length (for primary purpose of my extension).

Would it be possible to implement zero-vector padding on your extension too?
So we could compare, will it be better or not.

I propose putting the number "0" for padding token, since actually 0 stands for "!" while 256 stands for "!</w>" which is used each time the "!" parsed, so token #0 is impossible to enter in prompts anyway; using "0" in your extension the result is exactly the same as using "256".
So token 0 is not effective currently, and can be redefined to produce zero-filled vectors.

This will throw an AssertionError if you use prompt editing to have elements either appear or disappear.

A solution I found was to transition from/to _ instead of leaving it blank.

Traceback (most recent call last):
File "D:\SD\webui_forge_cu121_torch21\webui\modules\call_queue.py", line 57, in f
res = list(func(*args, **kwargs))
TypeError: 'NoneType' object is not iterable

when use it with sd1.5 modelm it works.

thx

This version of Auto1111 has the perfect features, save one, this extension. Any possibility of making it work? I was testing it and it does not function with it, although it does not make error messages or anything, it just doesn't function.

10k thanks!

I'm not sure if I am doing something wrong, but I feel like I've followed the steps correctly, apart from my prompt having 100 tokens.
But no matter how much I cut off, changed settings, tried batches of images, it seems like there is ALWAYS contamination instead.

analog style,model shoot style,RAW photo ((close portrait:1.6)) of an edgy and cool profile picture featuring a young man who resembles Sung Jin Woo from the popular webtoon series "Solo Leveling",
(trimmed` ,orange beard,:1.1),(green eyes:1.1),detailed eyes,
(black headphones:1.1),blue lightning around the headphones,solo focus,dressed in cool and fashionable attire,long coat, clean white backdrop background,feeling of edginess,sharpness,and coolness,capturing the essence of the character while presenting a visually striking profile picture,
lora:detailedeyes:0.9,
Negative prompt: Asian-Less-Neg,color contamination,
Steps: 50, Sampler: DPM++ SDE Karras, CFG scale: 10, Seed: 2077349957, Size: 512x728, Model hash: c194532de5, Model: realMoonAnime_v20, Cutoff enabled: True, Cutoff targets: ["orange", "green", "black", "blue"], Cutoff padding: _, Cutoff weight: 2.0, Cutoff disable_for_neg: True, Cutoff strong: False, Cutoff interpolation: lerp, Lora hashes: "detailedeyes: 3013f7bd7b29", Version: v1.3.1

The image above is the one with the cutoff, while the one below is without the cutoff.

Being honest, looks like there was more color contamination on the image with the cutoff activated than the one without it.

What exactly does Cutoff strongly do? More intensive color cut? Do I need to configure something for him? Does it have any side effects? Please tell me.

Hi. Thank you for developing nice extension.

I want settings of default target tokens.

I use this extension mainly for separating colors.
So, I always set target tokens as below.

red,blue,green,yellow,orange,aqua,white,black,pink,purple,brown,blonde

I want the above tokens to be default❗

Please consider it. Thanks.

Traceback (most recent call last):
  File "C:\stable-diffusion-webui\modules\call_queue.py", line 57, in f
    res = list(func(*args, **kwargs))
  File "C:\stable-diffusion-webui\modules\call_queue.py", line 36, in f
    res = func(*args, **kwargs)
  File "C:\stable-diffusion-webui\modules\txt2img.py", line 55, in txt2img
    processed = processing.process_images(p)
  File "C:\stable-diffusion-webui\modules\processing.py", line 734, in process_images
    res = process_images_inner(p)
  File "C:\stable-diffusion-webui\extensions\sd-webui-controlnet\scripts\batch_hijack.py", line 42, in processing_process_images_hijack
    return getattr(processing, '__controlnet_original_process_images_inner')(p, *args, **kwargs)
  File "C:\stable-diffusion-webui\modules\processing.py", line 858, in process_images_inner
    p.setup_conds()
  File "C:\stable-diffusion-webui\modules\processing.py", line 1314, in setup_conds
    super().setup_conds()
  File "C:\stable-diffusion-webui\modules\processing.py", line 469, in setup_conds
    self.uc = self.get_conds_with_caching(prompt_parser.get_learned_conditioning, negative_prompts, total_steps, [self.cached_uc], self.extra_network_data)
  File "C:\stable-diffusion-webui\modules\processing.py", line 455, in get_conds_with_caching
    cache[1] = function(shared.sd_model, required_prompts, steps, hires_steps, shared.opts.use_old_scheduling)
  File "C:\stable-diffusion-webui\modules\prompt_parser.py", line 188, in get_learned_conditioning
    conds = model.get_learned_conditioning(texts)
  File "C:\stable-diffusion-webui\modules\sd_models_xl.py", line 31, in get_learned_conditioning
    c = self.conditioner(sdxl_conds, force_zero_embeddings=['txt'] if force_zero_negative_prompt else [])
  File "C:\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1547, in _call_impl
    hook_result = hook(self, args, result)
  File "C:\stable-diffusion-webui\extensions\sd-webui-cutoff\scripts\cutoff.py", line 94, in hook
    assert isinstance(mod, CLIP)
AssertionError

Prompt: A man wearing a red shirt next to a blue space monkey
Negatives: Text, 3d

Cutoff tokens: red,blue

A1111 1.6

Happens with any prompt, not just this one.

I see that Cutoff saves settings in PNG Info of images:

When I click "send to ..." button in PNG Info tab, I expect those fields to be automatically set in the Cutoff section, but they are not. So I need to set them myself everytime.

I don't know if it is a bug in my stable diffusion setup, or there are not such a feature yet.

Thank you for the extension anyways.

Let me explain. As I understand, you:

1. Split prompt at each comma, if there is a target term between.
2. For each group create two its versions: original and replaced with target token.
3. Transform everything several times with CLIP.
4. Somehow combine target arrays so that each group mainly receives its original part while "sees" other groups as replaced.

What if we do this not only for the last clip layer (known as ClipSkip 1), but for each layer available? (Yes, it will work only for SD1 in this case, but it's worth trying!)

I propose something like this:

1. Replace every term, send to CLIP and extract each value at each layer (10x77x768, or so).
2. For each group, freeze all CLIP weights except for the current group, for all layers.
3. When CLIP will transform each part, it should not rewrite any frozen number which belongs to other groups (this can be imagined as "inpainting" all the way from the first to the last layer but only for small subset of vector embeddings).

(Personally, I don't know how technically hook/replace CLIP's behavior, but theoretically it should be possible).

In this scenario, there would be no single bit of color information leave its group! Though, the composition might change severely (closely resembling that with already replaced terms), and the colors may not play nicely with each other (or being washed-out), but we need to see it ourselves.

What do you think?