Open source
Week of 2 November 2026
This week's picks center on reducing inference steps and simplifying fine-tuning workflows for image and video diffusion models.
- 1Pika H3 FlexHF Models
Pika-Labs· 348 likes
Pika-H3-Flex is an image-to-video diffusion model built on MiniMax-H3 that generates video clips in six steps at native 480p or 768p, with upscaling to 4K available through the Pika API. Builders working on rapid iteration pipelines may find the six-step generation particularly useful for reducing inference time without switching to a distilled model of a different architecture.
diffuserssafetensorsvideo-generationaudio-video-generationloraimage-to-video - 2
op7418· 7,929 stars
A Claude Code / Codex agent skill that generates Xiaohongshu carousels and WeChat cover images by rendering single-file HTML templates to PNG via Playwright, with 28 layouts and 10 themes drawn from editorial and Swiss design systems. Builders working on agentic image generation pipelines may find the structured layout-plus-theme architecture a useful reference for how to expose design variation as parameters a coding agent can control directly.
agent-skillai-agentanthropicclaude-codeclaude-skillcodex - 3Pika H3 Flex Turbo LoRAHF Models
akhaliq· 72 likes
Pika-H3-Flex-Turbo-LoRA is an image-to-video model that fuses Pika's motion style with few-step distillation into standard LoRA files for MiniMax-H3, supporting 2-step, 4-step, and 8-step inference modes. Builders looking to reduce generation time without switching pipelines may find it useful, as it loads as a single file onto stock H3 setups in ComfyUI or Diffusers without requiring custom nodes or model patches.
minimax-h3pikapika-h3-flexturbo4-step8-step - 4Anima Ref AdapterHF Models
ElectronGirl· 55 likes
A LoRA adapter for the Anima anime model that extracts character identity (hair, eyes, outfit, accessories) from a single reference image while letting a text prompt independently control scene, pose, and framing. Builders working on anime pipelines who need consistent character appearance across varied compositions can integrate it directly into ComfyUI via dedicated custom nodes.
loraanimacomfyuireference-imagecharacter-consistencystyle-transfer - 5lora-scriptsGitHub
Akegarasu· 6,206 stars
A collection of Python scripts and a GUI for training LoRA and Dreambooth models on Stable Diffusion, built on top of kohya-ss's training framework. Builders looking to fine-tune diffusion models on custom subjects or styles without writing training code from scratch will find the included GUI a practical starting point.
dreamboothfinetunelorastable-diffusion