diffusion

NVIDIA releases Nemotron-Labs-Diffusion, the first tri-mode language model family (3B/8B/14B) that switches between autoregressive, diffusion, and self-speculation decoding by changing the attention pattern, achieving up to 4× higher real throughput.

NVIDIA released the Nemotron-Labs-Diffusion model family (3B to 14B) that supports both AR and diffusion decoding with novel self-speculation, achieving significant speedups (up to 4x) over standard AR and Eagle3 methods across hardware platforms.

Open-dLLM adapts Qwen3.6 to use diffusion-based generation, achieving over 3,000 tok/s on an RTX 5090 for short sequences, with code released on GitHub.

Orthrus is a dual-architecture framework that combines autoregressive LLM fidelity with diffusion model speed, delivering up to 7.8x speedup on Qwen3 models while guaranteeing identical output distribution.

Introduces Orthrus, a method that injects a trainable diffusion attention module into a frozen autoregressive transformer to achieve up to 7.8× tokens per forward pass and ~6× wall-clock speedup on MATH-500, with provably identical output distribution to the base Qwen3-8B model. The approach requires minimal additional parameters and training, and avoids the TTFT penalty of external drafters.

ByteDance releases Cola-DLM, a hierarchical continuous latent-space diffusion language model combining a Text VAE with a block-causal Diffusion Transformer, available on Hugging Face with model weights, code, and paper.

Microsoft releases Lens, a 3.8B-parameter foundational text-to-image model with efficient training and fast high-resolution generation, featuring dense-caption pre-training and mixed-resolution learning.

This paper introduces DiHAL, a diffusion-transformer hybrid that uses geometry-based proxies to select a layer in a pretrained language model for hidden-state replacement with a diffusion bridge, improving continuous diffusion language modeling by avoiding direct token recovery.

Microsoft releases Lens, a 3.8B-parameter foundational text-to-image model designed for efficient training and fast high-resolution generation, achieving competitive quality with reduced compute.

Introduces Reference-Guided Flow Matching, a method that uses a reference distribution to guide the flow matching process, improving sample quality and generation efficiency.

PRISM is a diffusion-based framework for text image super-resolution that uses flow-matching prior rectification and uncertainty-aware residual encoding to improve accuracy under severe degradation, achieving state-of-the-art performance with millisecond-level inference.

This Hugging Face model page introduces an IC-LoRA trained on top of LTX-2.3-22b for lip dubbing, with a project page, paper, and inference pipeline available.

This paper introduces BLT Diffusion and speculative decoding techniques for byte-level language models to significantly reduce generation latency and memory bandwidth costs while maintaining quality.

LLaDA2.0-Uni unifies multimodal understanding and generation within a single diffusion-based large language model architecture.

unsloth releases a GGUF quantized version of Baidu's ERNIE-Image-Turbo model using Unsloth Dynamic 2.0 methodology, enabling efficient text-to-image generation in 8 inference steps on consumer GPUs with 24GB VRAM.

Comfy-Org has repackaged Baidu's ERNIE-Image and ERNIE-Image-Turbo models for ComfyUI integration, providing ready-to-use model files organized for the ComfyUI node-based image generation framework.

Netflix releases VOID, a video inpainting model that removes objects from videos while realistically simulating physical interactions (e.g., objects falling when a person is removed), built on CogVideoX and fine-tuned with interaction-aware quadmask conditioning.

BFS: Best Face Swap is a collection of LoRAs for face, head, and body replacement across Qwen Image 2.1, Qwen Image Edit 2509/2511, Flux 2 Klein, Krea 2, and experimental LTX-2 workflows, with per-version guides and downloadable safetensors weights.

Black Forest Labs releases FLUX.1-dev, a 12-billion parameter open-weights text-to-image transformer model, available on Hugging Face with API endpoints and local inference support.

OpenAI presents improved techniques for training consistency models that enable high-quality single-step image generation without distillation, achieving significant FID improvements on CIFAR-10 and ImageNet 64×64 through novel loss functions and training strategies.