Jasper Research releases cookbook on building a text-to-image model from scratch, with 100M-image dataset and tiny model codebase

Developments·Benchmark

ImageBench releases 192-prompt T2I benchmark with 52 models and 9k images

A new text-to-image benchmark, ImageBench, evaluates 52 models on 192 curated difficult prompts covering text rendering, spatial reasoning, human realism, and negations. A VLM judges outputs against binary ground-truth questions, and all results and images are published publicly.

A new text-to-image benchmark, ImageBench, evaluates 52 models on 192 curated difficult prompts covering text rendering, spatial reasoning, human realism, and negations. A VLM judges outputs against binary ground-truth questions, and all results and images are published publicly.

Releases·Model Release

AuroraIMG-6M: 6M-Parameter Text-to-Image Model Generates Recognizable 64x64 Images

AuroraIMG-6M, a 6M-parameter text-to-image model, was released with model card and samples on Hugging Face. It produces recognizable images at 64x64 resolution with the parameter count covering the entire system, including no separate text encoder or VAE. Its small footprint makes it notable for practitioners exploring minimal text-to-image architectures.

AuroraIMG-6M, a 6M-parameter text-to-image model, was released with model card and samples on Hugging Face. It produces recognizable images at 64x64 resolution with the parameter count covering the entire system, including no separate text encoder or VAE. Its small footprint makes it notable for practitioners exploring minimal text-to-image architectures.

Releases·Model Release

SupraLabs releases Supra2-IMG, a 100M-parameter DiT text-to-image model trained from scratch on one H100

SupraLabs released Supra2-IMG, a 100M-parameter DiT text-to-image model trained from scratch in under 10 hours on a single H100 rented on Runpod. The open-weights model generates 256x256 images at claimed state-of-the-art quality and runs locally, producing an image in roughly 20 seconds on CPU or 2 seconds on GPU. It is available on Hugging Face with an inference script.

SupraLabs released Supra2-IMG, a 100M-parameter DiT text-to-image model trained from scratch in under 10 hours on a single H100 rented on Runpod. The open-weights model generates 256x256 images at claimed state-of-the-art quality and runs locally, producing an image in roughly 20 seconds on CPU or 2 seconds on GPU. It is available on Hugging Face with an inference script.

A new paper introduces Concept Guidance, a training-free method for precise, concept-specific control in text-to-image diffusion models. It addresses two major drawbacks: lack of continuous guidance for concepts like aesthetics and unreliability for tasks requiring local coherence such as text and hands. The method leverages concept-wise mutual information and layer-specific differences to enable targeted generation.

A new architecture called Giraffe translates hidden text representations from multimodal LLMs into visual embeddings for graphic design generation, addressing the input-length overhead of existing multi-token image methods.