Open-Source GPT-4o Image Editor: BAGEL & AI News

TL;DR

ByteDance released BAGEL, a free open-source multimodal model that generates and edits images, understands visuals, and renders text accurately, with all models available on HuggingFace and GitHub. The same week brought Anthropic's Claude 4 for coding, Google I/O AI updates, and a study finding leading AI models scored 81% on EQ tests versus 56% for humans.

Transcript

AI never sleeps. And this week has been absolutely insane. We finally have an open-source GPT40 image generator and editor. This open- source AI is really good at medical analysis. We have an AI that's going to kill my job. It can make full explainer videos. Anthropic released their long- aaited model Claude 4, and it seems to be really good at cod... Read More

Key Insights

• BAGEL by ByteDance is an open-source multimodal language model similar to GPT-4o that can chat, generate images, edit images, and answer questions about images, with all models released on HuggingFace and their GitHub repo for free offline use.
• BAGEL's visual understanding lets it identify people in blurry photos, solve for X from an uploaded image, give cooking instructions from a photo of a dish, and identify movies like The Godfather from a single frame.
• BAGEL renders text inside images accurately, correctly labeling three antique potions as XDXL, bagel, and flux, which many image generators struggle to do.
• BAGEL includes a thinking feature like Deepseek R1, letting it reason step-by-step before responding, such as deducing that an animal with 'seven plus two lives' means a cat with nine lives.
• BAGEL edits existing images while keeping unchanged parts consistent, supporting object replacement, watermark removal, virtual clothing try-on, style transfer to anime or Ghibli, and simulated camera movement frame by frame.
• MTVCrafter transfers movement from a reference video onto any reference character by converting the video into 3D motion data, tokenizing it through a 4D motion tokenizer, and mapping movements via a motion video model.
• A Nature study tested six AI models including GPT-4, GPT-01, Gemini 1.5 Flash, Copilot 365, Claude 3.5 Haiku, and Deepseek v3, finding they averaged 81% on standard EQ tests versus 56% for human participants.
• The week's releases included Anthropic's Claude 4, described as good at coding, Google's I/O AI updates, MedGemma for medical analysis, LearnLM, and Microsoft's open-source NLWeb and GitHub Copilot coding agents.

Questions & Answers

Q: What is BAGEL by ByteDance?

BAGEL is an open-source multimodal language model from ByteDance, similar to GPT-4o. You can talk to it like a chatbot, prompt it to generate images, edit existing images, or ask it questions about images, giving it visual understanding. ByteDance released all the models already, available on HuggingFace and their GitHub repo, so you can download and use it offline for free.

Q: How to download and use BAGEL for free?

All of BAGEL's models are available on HuggingFace plus on ByteDance's GitHub repo, which contains all the instructions on how to download and use it on your computer offline for free. The presenter notes that VRAM requirements are unfortunately not specified anywhere in the documentation. Links to the main BAGEL page are provided in the video description for anyone interested in reading further.

Q: What image editing capabilities does BAGEL have?

BAGEL can edit existing images in many ways while keeping the rest of the photo consistent with the original. It can add objects like a rabbit for someone to hold, swap clothes for virtual try-on, replace objects such as swapping a dog for a boy, remove watermarks, apply style transfer into 3D animated, Japanese anime, clay, or Ghibli styles, and simulate camera rotation or forward movement frame by frame.

Q: Can BAGEL render text accurately inside generated images?

Yes, BAGEL is very good at rendering text in images. In one example, prompted to create three antique glass magic potions labeled XDXL, bagel, and flux respectively, it produced exactly that with correct labels. The presenter notes this is something many image generators cannot get correct, alongside BAGEL's strong prompt understanding for tricky requests like a cat facing a mirror showing a snow leopard reflection.

Q: How does BAGEL's thinking feature work?

BAGEL can think through its answer like Deepseek R1 and most other state-of-the-art models before giving a response, allowing it to ponder longer and give better answers. For example, asked to generate a prompt for an animal with 'seven plus two lives,' it reasons step-by-step to figure out the user wants nine lives, which popular culture links to cats, then generates a prompt producing a cat surrounded by nine heart symbols.

Q: What is MTVCrafter and how does it work?

MTVCrafter is an open-source AI that takes a reference photo of any character plus a reference video of someone moving, and transfers those movements onto the reference character. It works with flat vector art, multiple characters in sync, and 3D animated characters. The reference video is first converted into 3D motion data, passed through a 4D motion tokenizer that breaks motion into smaller chunks, then fed through a motion video model to map the movements.

Q: Do AI models have higher emotional intelligence than humans?

According to a study cited in the video, yes. Researchers tested six leading AI models including GPT-4, GPT-01, Gemini 1.5 Flash, Copilot 365, Claude 3.5 Haiku, and Deepseek v3, and found they achieved an average score of 81% on standard EQ tests, compared to only 56% for human participants. The presenter notes these were older models, yet they still outperformed humans at selecting emotionally intelligent answers.

Q: What other AI updates were announced this week?

Alongside BAGEL and MTVCrafter, the week featured Anthropic's long-awaited Claude 4, which seems good at coding, and Google's massive AI updates during their I/O event including video overviews, MedGemma for medical analysis, and LearnLM with quizzes. Microsoft also released open-source tools including NLWeb, which brings conversational interfaces directly to the web, and GitHub Copilot coding agents, as noted in the video.

Summary & Key Takeaways

• ByteDance released BAGEL, an open-source GPT-4o-style multimodal model available free on HuggingFace and GitHub. It chats, understands images, generates images with accurate text rendering, and edits photos through object replacement, style transfer, virtual try-on, and simulated camera movement, plus a step-by-step reasoning feature.

• MTVCrafter transfers movements from a reference video onto any reference character, including flat vector art, multiple characters, and 3D animated figures. It converts video into 3D motion data, tokenizes it, then maps movements, though the presenter notes the output quality lags behind tools like Alibaba's Vace.

• A Nature study found six AI models averaged 81% on emotional intelligence tests versus 56% for humans, despite using older models. The week also featured Claude 4 for coding, Google I/O updates, MedGemma, LearnLM, and Microsoft's open-source NLWeb and GitHub Copilot agents.