AI Art Weekly #119

Hello, my fellow dreamers, and welcome to issue #119 of AI Art Weekly! 👋

After last week’s Gemini 2 multimodality, ChatGPT finally released their GPT-4o native image generation feature which they had presented almost over a year ago.

GPT-4o, the model itself, can now generate images. This is not to be confused with DALL-E, a separate image model which GPT-4o only called in the background up until now.

And oh boy, is it good. Slow and heavily restricted, but very good. It instantly caused a “ghiblify” trend on X after launch. Never before have I seen the entire feed just filled with a single trend before.

4o is capable of so much more though. Making comics, movie posters, YouTube thumbnails, turning things real, transparent illustrations, ads, and basically any image task one can imagine that so far required separate models, workflows, and pipelines - most of them are now just one prompt away.

Buckle up, because this is the power of multimodality.

---

News & Papers

3D

DIDiffGes can generate high-quality gestures from speech in just 10 sampling steps.

PGC: Physics-Based Gaussian Cloth from a Single Pose can create simulation-ready garments from a single static pose. It captures fine details using 3D Gaussian splats and allows accurate garment simulation on new poses with minimal data.

PhysGen3D can turn a single image into an interactive 3D scene. It allows users to control object speed and material properties while ensuring the scene looks realistic and behaves like it would in real life.

Liv3Stroke can reconstruct dynamic 3D sketches from videos using flexible 3D strokes.

LoRA-MDM can generate stylized human motions in different styles, like “Chicken,” by using a few reference samples with a motion diffusion model. It allows for style blending and motion editing while keeping a good balance between text fidelity and style consistency.

DeClotH can reconstruct 3D cloth and human bodies from a single image.

LookCloser can capture the full structure and fine details of scenes in one NeRF model. It allows for real-time rendering and interaction.

TaoAvatar can create high-quality, real-time full-body talking avatars from multiple camera angles. It runs at 90 frames per second on devices like the Apple Vision Pro and allows detailed control over facial expressions and body movements, making it great for e-commerce live streaming and holographic communication.

SV4D 2.0 can generate high-quality 4D models and videos from a reference video.

DiffPortrait360 can create high-quality 360-degree views of human heads from single images.

Image

LeX-Art can generate high-quality text-image pairs with better text rendering and design. It uses a prompt enrichment model called LeX-Enhancer and two optimized models, LeX-FLUX and LeX-Lumina, to improve color, position, and font accuracy.

SISO can generate and edit images using just one subject image without any training. It improves image quality, keeps the subject clear, and preserves the background better than other methods.

MagicColor can automatically colorize multi-instance sketches while keeping colors consistent across objects using reference images.

Video

AccVideo can speed up video diffusion models by reducing the number of steps needed for video creation. It achieves an 8.5x faster generation speed compared to HunyuanVideo, producing high-quality videos at 720x1280 resolution and 24fps, which makes text-to-video generation way more efficient.

PP-VCtrl can turn text-to-video models into customizable video generators. It uses control signals like Canny edges and segmentation masks to improve video quality and control without retraining the models, making it great for character animation and video editing.

Video Motion Graphs can generate realistic human motion videos by combining clips from a reference video with music or motion tags. It uses HMInterp for smooth transitions and high-quality video, improving how we create multi-modal human motion videos. Pretty smart.

Synthetic Video Enhances Physical Fidelity in Video Synthesis can improve the realism of video generation by using synthetic videos that follow real-world physics. This method reduces unwanted artifacts and ensures 3D consistency, making generated videos more lifelike.

FullDiT can generate videos with control over camera angles, identities, depth, and text. It uses a unified full-attention mechanism to improve video content creation and achieve top results in multi-task video generation.

Mask²DiT can generate long videos with multiple scenes by aligning video segments with text descriptions.

HunyuanPortrait can animate characters from a single portrait image by using facial expressions and head poses from video clips. It achieves lifelike animations with high consistency and control, effectively separating appearance and motion.

DisentTalk can generate high-quality talking face animations from audio. It ensures accurate lip synchronization and natural expressions while maintaining smooth movement over time.

Audio

MusiCoT can generate high-quality music by first outlining its structure and then producing audio tokens. It supports variable-length audio inputs and performs well in both objective and subjective measures, effectively addressing copying issues.

Also interesting

There is just so much 4o can do, so here are a few more examples for inspiration.

And that my fellow dreamers, concludes yet another AI Art weekly issue. If you like what I do, you can support me by:

• Sharing it 🙏❤️
• Following me on Twitter: @dreamingtulpa
• Buying me a coffee (I could seriously use it, putting these issues together takes me 8-12 hours every Friday 😅)
• Buying my Midjourney prompt collection on PROMPTCACHE 🚀
• Buying access to AI Art Weekly Premium 👑

Reply to this email if you have any feedback or ideas for this newsletter.

Thanks for reading and talk to you next week!

– dreamingtulpa