zhanghaotian0225

zhanghaotian0225 / Accumulative-Decoding

Public

Mitigating Hallucinations in Large Vision-Language Models via Accumulative Decoding

19
6
100% credibility
Found Mar 27, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This project offers a lightweight, no-training-needed technique to reduce false details in AI-generated image descriptions using vision-language models like LLaVA, including scripts to evaluate performance on key benchmarks.

How It Works

1
🔍 Discover the tool

You hear about Accumulative Decoding, a simple way to make AI assistants describe images more accurately without making up details.

2
📥 Gather your setup

Download the tool and prepare a ready-to-use AI assistant specialized in understanding pictures.

3
🖼️ Pick an image

Choose a photo or picture you want the AI to analyze, like a scene or object.

4
✨ Turn on accurate mode

Activate the special feature that keeps the AI focused on what's really in the image throughout its entire response.

5
💬 Ask about the image

Give the AI a question or request to describe the picture in detail.

6
📊 Test on challenges

Run quick checks on standard image-understanding tests to see the improvements.

🏆 Enjoy reliable results

Your AI now gives spot-on descriptions without hallucinations, beating benchmarks and feeling trustworthy.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is Accumulative-Decoding?

Accumulative-Decoding is a Python library that mitigates hallucinations in large vision-language models by injecting a cumulative visual-grounding signal into every step of autoregressive text generation. It tackles the problem of models confidently outputting details not present in images, like in LLaVA-1.5-7B, without any training or model changes—just plug it into HuggingFace's generate() as a logits processor. Developers get more faithful image descriptions and answers on benchmarks like MME, MM-Vet, and MMMU.

Why is it gaining traction?

It stands out for being fully training-free and self-reinforcing, outperforming baselines like DoLA and VCD on perception tasks while preserving fluency—no hyperparameter tuning needed beyond defaults. The hook is its drop-in simplicity: extract a visual embedding once per image, pass the processor, and watch hallucination rates drop across long outputs. Python/PyTorch users appreciate the eval scripts for quick benchmarking on multimodal datasets.

Who should use this?

AI engineers building vision-language apps, like visual question answering or image captioning tools, who need reliable outputs from open models without fine-tuning costs. It's ideal for researchers validating LLaVA-style models on hallucination-heavy tasks such as OCR, spatial reasoning, or existence detection in production prototypes.

Verdict

Grab it if you're iterating on LVLMs and want an easy win on hallucination mitigation—docs and setup are solid for a 19-star repo. At 1.0% credibility from low adoption, treat it as experimental; test thoroughly before deploying.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.