clip-flant5-xxl

CLIP-FlanT5-XXL (VQAScore) — trust_remote_code HuggingFace format

Unofficial transformers-format port of CLIP-FlanT5-XXL, the VQAScore judge from "Evaluating Text-to-Visual Generation with Image-to-Text Generation" (Lin et al., 2024). It bundles a self-contained modeling_clipflant5.py and an auto_map, so the model loads with AutoModelForSeq2SeqLM.from_pretrained(..., trust_remote_code=True) under transformers 5.X (and 4.57) with no llava / t2v_metrics dependency and no pinned old transformers.

Not an official release. Model, weights, and architecture (CLIPT5ForConditionalGeneration) are by Zhiqiu Lin et al. — original weights zhiqiulin/clip-flant5-xxl (Apache-2.0), code https://github.com/linzhiqiu/CLIP-FlanT5 and https://github.com/linzhiqiu/t2v_metrics, paper arXiv:2404.01291. This repo re-packages that checkpoint for trust_remote_code loading and is used as the paper-strict VQAScore judge for the CVTG-2K benchmark (TextCrafter, arXiv:2503.23461).

What it is

A CLIP ViT-L/14-336 vision tower + a FlanT5-XXL encoder-decoder. Image patch features are projected and spliced into the T5 encoder inputs at an <image> placeholder; the T5 decoder then scores an answer. VQAScore = the probability the decoder assigns to "Yes" for the question 'Does this figure show "{caption}"? Please answer yes or no.' (exp(-CrossEntropy) under teacher forcing) — a reference-free image↔text alignment score.

modeling_clipflant5.py consolidates the upstream clip_t5.py + vision tower + projector + preprocessing (t5_tokenizer_image_token, expand2square) into one module using only public T5/CLIP transformers API.

transformers-5.X note. generate() / prepare_inputs_for_generation are intentionally removed — they indexed the legacy tuple KV-cache (gone in 5.X), and VQAScore needs only the teacher-forced forward() + logits. Use this model for scoring, not free-form generation.

Usage (scoring)

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("bigshanedogg/clip-flant5-xxl", trust_remote_code=True).eval()
tokenizer = AutoTokenizer.from_pretrained("bigshanedogg/clip-flant5-xxl")
# Preprocess the image with model.vision_tower.image_processor, wrap the question with
# `format_question(...)` (t5_chat), tokenize with `t5_tokenizer_image_token`, then run a teacher-forced
# forward with labels="Yes" and read exp(-CrossEntropy) — see modeling_clipflant5.py's docstring.

Licensing

• CLIP-FlanT5 model / weights / code (Zhiqiu Lin et al.): Apache-2.0 — https://github.com/linzhiqiu/CLIP-FlanT5
• FlanT5-XXL base (Google): Apache-2.0. OpenAI CLIP ViT-L/14-336 vision tower: MIT.
• This port (modeling_clipflant5.py, packaging): Apache-2.0 — see LICENSE; attribution in NOTICE.