Open source
Week of 5 October 2026
This week's picks center on drop-in adapters that extend existing image and video pipelines without requiring retraining.
- 1BFS Best Face SwapHF Models
Alissonerdx· 1,198 likes
BFS is a collection of LoRA weights for face, head, and body swapping built across several base models including Qwen Image 2.1, Qwen Image Edit, Flux 2 Klein, and experimentally LTX-2, with each variant shipping its own dedicated workflow files and prompt contracts. Builders targeting a specific pipeline can drop in the corresponding safetensors without adapting a one-size-fits-all approach, which makes it practical for integrating precise identity replacement into existing diffusers-based workflows.
diffusersloraqwen-imageqwen-image-2.1qwen-image-editface-swap - 2bragGitHub
latent-spaces· 13,658 stars
`latent-spaces/brag` wraps a single command around the process of converting a finished project into a short launch video. Builders who want to skip the manual work of demoing and packaging their work for sharing may find it worth a look, particularly given its integration with Claude-based agent tooling.
agent-skillsclaude-codeclaude-skillshyperframeslaunch-videovideo-generation - 3QI_2.1_AnyAngleHF Models
lilylilith· 133 likes
A LoRA adapter for Qwen Image 2.1 that enables camera angle re-rendering of an existing image by using a Gaussian splat or 3D model (generated via tools like Tripo or Trellis) as a coarse geometric reference, then refining the result to preserve the original image's style. Builders who need precise, arbitrary viewpoint control rather than fixed azimuth/elevation presets may find this pipeline worth integrating into multi-step image editing workflows.
LoRAloraQwen-Image-2.1image-to-imagebase_model:Qwen/Qwen-Image-2.1base_model:adapter:Qwen/Qwen-Image-2.1 - 4MiniMax H3 X2 Detail VAEHF Models
speach1sdef178· 81 likes
A fine-tuned VAE for MiniMax H3 that decodes video latents at 2x spatial resolution while offering an optional ComfyUI detail-enhancement node as a second operating mode from the same weights. Builders already running MiniMax H3 pipelines in ComfyUI can drop it in as a direct VAE replacement without retraining, making it a low-friction path to higher-resolution video output.
minimax-h3minimaxcomfyuivaevideo-vaevideo-generation - 5Qwen Image 2.1 Consistency LoRAHF Models
ausboss· 76 likes
A LoRA for Qwen Image 2.1 that reduces the spatial drift and unintended repainting the base model introduces during image edits, helping outputs stay geometrically aligned with the source. If pixel-level consistency across seeds matters for your editing pipeline, it ships with two ComfyUI workflows covering both constrained edits and a fallback realignment approach for edits that require subject movement.
loraimage-editingconsistencyqwen-imageqwen-image-2.1comfyui