Open source

This week's picks focus on reducing the cost of high-quality video and image generation through distillation, efficient architectures, and modular pipelines.

• 1

  LTX Best Face IDHF Models

  Alissonerdx· 118 likes

  A reference-to-video LoRA built on LTX-2.3 (22B) that takes a photo of a person and a text prompt to generate identity-consistent video, using overlap reference conditioning, TASS-RoPE, and an ArcFace identity loss during training. Builders working in ComfyUI who need per-subject identity preservation without per-video fine-tuning may find the dual reference modes (close-up face and full character sheet with clothing consistency) worth evaluating.

  ltx-videoidentity-preservationipt2vreference-to-videoloracomfyui
• 2

  Lingbot World V2 14b Causal FastHF Models

  robbyant· 90 likes

  A 14B-parameter image-to-video world model built on a causal pretraining paradigm, **lingbot-world-v2-14b-causal-fast** generates open-ended interactive video from a single image with an unbounded temporal horizon. Builders working on real-time video synthesis pipelines may find the distilled "fast" variant worth benchmarking, as the model card claims it can drive 720p video streams at 60 fps.

  diffuserssafetensorsi2vWorld Modelimage-to-videoen
• 3

  ViMaxGitHub

  HKUDS· 11,118 stars

  ViMax frames video generation as a multi-agent pipeline where distinct roles (director, screenwriter, producer, and generator) collaborate to produce videos from high-level prompts. Builders exploring agentic approaches to long-form or structured video synthesis may find the role decomposition architecture a useful reference for breaking down complex generation tasks.

  agentic-aigcvideo-generation
• 4

  M87HF Models

  mgwr· 77 likes

  M87 is an aesthetic LoRA adapter for KREA-2 Turbo, trained on a 100-image curated dataset to improve compositional quality, lighting, and visual texture across generations. Builders looking to add a consistent art-directed quality to KREA-2 Turbo outputs without constraining subject matter may find it a useful drop-in enhancement layer.

  diffuserstext-to-imageloratemplate:diffusion-lorabase_model:krea/Krea-2-Turbobase_model:adapter:krea/Krea-2-Turbo
• 5

  SanaGitHub

  NVlabs· 8,472 stars

  NVlabs/Sana is a text-to-image generation system built around a linear diffusion transformer architecture designed for efficient high-resolution image synthesis. Builders looking to reduce inference costs without abandoning diffusion-based quality will find the linear attention approach worth studying as an alternative to standard quadratic-complexity transformers.

  diffusionditlinear-transformernvfp4pytorchreinforcement-learning