Open source
Week of 12 October 2026
Quantized and optimized image and video generation models are making local inference more accessible across varied hardware setups.
- 1Qwen Image 2.1 Uncensored QuantizedHF Models
abenzerps· 4,070 likes
Qwen-Image-2.1-Uncensored-Quantized packages Alibaba's Qwen-Image-2.1 text-to-image model in multiple quantized formats, including GGUF, Safetensors, and MLX variants ranging from BF16 down to 4-bit, enabling local inference across a range of hardware budgets. Builders targeting ComfyUI workflows or Apple Silicon will find ready-to-use files without needing to run their own quantization pipeline.
ggufqwenimage-generationcomfyuicomfyui-ggufmlx - 2Qwen Image 2.1 TurboHF Models
Qwen· 610 likes
Qwen-Image-2.1-Turbo is a 7B text-to-image and image-editing model that completes generation in just 8 denoising steps using a bundled sampling schedule and prefix KV caching to reuse text and reference-image context across steps. Builders looking to reduce inference latency without reconfiguring a custom scheduler may find the drop-in `QwenImage21Pipeline` integration with Diffusers a practical starting point.
diffuserssafetensorsqwenimage-generationimage-editingtext-to-image - 3HunyuanVideoGitHub
Tencent-Hunyuan· 12,614 stars
HunyuanVideo is Tencent's open-source framework for large-scale video generation using diffusion transformers. Builders working on text-to-video pipelines may find it worth examining for its systematic approach to scaling diffusion-based architectures, backed by a codebase that has drawn significant community traction.
diffusion-modelsdiffusion-transformervideo-generation - 4Qwen Image 2.1 Turbo Uncensored GGUFHF Models
AtomicChat· 364 likes
A GGUF-quantized build of Qwen3-VL-8B-Instruct serving as the text encoder component for the Qwen-Image-2.1-Turbo pipeline, designed as a drop-in `--llm` argument for stable-diffusion.cpp. Builders running local inference stacks with stable-diffusion.cpp can slot it directly into an existing pipeline without changes to the denoiser or surrounding tooling.
ggufatomic-chatqwenqwen-imageqwen3-vltext-encoder - 5Iris 3bHF Models
speridlabs· 185 likes
Iris-3B is a 3-billion-parameter flow-matching model that generates images directly in pixel space without a VAE or latent bottleneck, and can be fine-tuned for depth estimation and image restoration and upscaling. Builders exploring pixel-space diffusion as a vision foundation alternative to models like DINOv2 may find it worth examining for downstream perception tasks.
pytorchsafetensorstext-to-imagediffusionpixel-spaceflow-matching