README.md

CLIP-Interrogator-LongCLIP-hallucinwords

CLIP Interrogator, fully in HuggingFace Transformers 🤗, with LongCLIP & CLIP's own words and / or your own words!

• 🔎 Why? Because even in 2025, CLIP is still SOTA for t2i / t2v models (as a Text Encoder)!
• Refactored to use HuggingFace Transformers ✨, allows easy loading of custom CLIP models.
• Also supports Long-CLIP models with 248 tokens. And basically any other CLIP. ✨

😊👍

• 📄 You can insert own words with --ownwords. Will use data/ownwords.txt (put your file there!).
• You can save the results as .csv, or as .txt next to images with {image_filename}.txt. Or both.
• 📁 --output, choices=['rename', 'csv', 'txt', 'both']. For example, to save both: --output both
• 📉 Reduce memory use (especially for LongCLIP and / or with BLIP-2), e.g.: --chunk_size 1024 (default 2048)
• Use python clip-hallucin-interrogator.py --help (or clip-classic-interrogator.py) for more options.

🤔⚠️

• clip-classic-interrogator.py is like the original pharmapsychotic/clip-interrogator, just for transformers.
• clip-hallucin-interrogator.py uses the original + in addition, CLIP's own words, obtained in gradient ascent for many diverse images.
• ⚠️ The 'hallucin' words are NOT filtered. Will very likely contain sensitive / offensive / NSFW words that may be produced even if you have PG-13 images.
• ⚠️ Contains 'sneaky' offensive words that cannot be detected by simple NLP. Hence I cannot confidently provide a separate (safe, filtered) version.
• 🔎 A real example CLIP word: aggravfckremove - is: aggravated + look again: aggravfckremove. Yes. It's a CLIP concept of 'being angry and violent'.
• ⚠️ Alas, clip-hallucin-interrogator.py is for research / personal (and responsible) use only.
• Remember that you can use --mode negative (choices=['best', 'classic', 'fast', 'negative']) to steer away from unwanted concepts with a negative prompt.
• Consider replacing data/wCLIPy_negative.txt with a "blacklist words github" (just google it). But still, don't use this for publicly accessible stuff.
• 👉 Also remember: Every chosen word is the best match for the image (according to CLIP, at least; and CLIP guides your generative model).

☝️🤓

# Example usage: 

# Fine-tuned SAE-CLIP (huggingface.co/zer0int), including additional 'trippywords':
python clip-hallucin-interrogator.py --output csv --outfile saeclipwords --image_folder images --m_clip zer0int/CLIP-SAE-ViT-L-14 --trippywords

# LongCLIP with a smaller batch size:
python clip-hallucin-interrogator.py --output csv --outfile longclipwords --image_folder images --m_clip zer0int/LongCLIP-SAE-ViT-L-14 --chunk_size 1024

# Fine-tuned GmP-CLIP with BLIP-2, save both .txt and .csv:
python clip-hallucin-interrogator.py --output both --outfile gmpblip2 --image_folder images --m_clip zer0int/CLIP-GmP-ViT-L-14 --m_caption blip2-2.7b

# Fine-tuned LongGmP-CLIP with BLIP-2 in 'fast' mode:
python clip-hallucin-interrogator.py --output csv --outfile longgmpblip2 --image_folder images --m_clip zer0int/CLIP-GmP-ViT-L-14 --m_caption blip2-2.7b --mode fast

# See all options!
python clip-hallucin-interrogator.py --help

👨‍💻🤖

• DIY wordlist using CLIP gradient ascent (yes, all above caveats apply ⚠️):
• 🔎 Why? CLIP knows best what a CLIP sees (and will subsequently guide a diffusion model into).
• But gradient ascent is expensive (compute). Give it a few representative images, get a CLIP 'opinion', re-use the words in CLIP-Interrogator for all images!✨
• Usage: python diy-0-run-gradient-ascent.py --img_folder path/to/myimages (or --img_folder images as example)
• Then: python diy-1-preprocess-words.py. Result: a .txt file with all the words.
• Clean them (manually review, delete weird ones), and replace data/ownwords.txt with the file.
• 🕵️
• Optional: If you gave CLIP a lot of images, you'll likely have an overwhelming amount of words:
• Use python diy-2-make-clusters-DBSCAN.py to leverage a clustering algorithm to sort them out.
• Clustering with DBSCAN will do 80% of the job, but you'll still have to categorize and review the individual clusters manually.
• [Or use the OpenAI API and ask GPT-4o or a similar SOTA model]. Ain't no simple NLP gonna understand CLIP's crazy words!
• Example for CLIP words belonging to (I guess) an 'animals' concept cluster found with DBSCAN:

• The script also saves plots. Check them. Very scattered clusters? Likely noise words. Is cluster AND seems to include your concept (even if in a weird way)? Valid words.
• Later clusters (higher number) are more likely to contain noise, e.g. 'asdfghj' (yes, that's a token in CLIP). Remove suspicious / unclear words.
• Edit python diy-1-preprocess-words.py to point to the folder with the edited clusters, save to data/ownwords.txt
• Remember to use the argument --ownwords to include your DIY words.
• If they don't show up in the result, they were worse than the other choices in CLIP Interrogator / didn't generalize to all of your images. Try again using more images.

✔️

# Change caption model, for example:
--m_caption blip2-2.7b

blip-base: 'Salesforce/blip-image-captioning-base',   # 990MB
blip-large: 'Salesforce/blip-image-captioning-large', # 1.9GB
blip2-2.7b: 'Salesforce/blip2-opt-2.7b',              # 15.5GB
blip2-flan-t5-xl 'Salesforce/blip2-flan-t5-xl',      # 15.77GB
git-large-coco: 'microsoft/git-large-coco',           # 1.58GB

# Usage to change CLIP model (use any from HF):

--m_clip openai/clip-vit-large-patch14
--m_clip zer0int/CLIP-GmP-ViT-L-14
--m_clip zer0int/CLIP-SAE-ViT-L-14
--m_clip zer0int/LongCLIP-L-Diffusers
--m_clip zer0int/LongCLIP-GmP-ViT-L-14
--m_clip zer0int/LongCLIP-SAE-ViT-L-14