Open source
Week of 7 September 2026
MiniMax-H3 adaptations dominate this week, from LoRA controllers to fused checkpoints and faster inference architectures.
- 1Vdn Minimax H3HF Models
OpenVDN· 218 likes
VDN-Minimax-H3 is a text-to-video diffusion model built on MiniMax H3 that uses a hybrid attention architecture, combining a frame-wise linear attention branch with a softmax branch, to generate video faster than real-time playback on 8 B200 GPUs. Builders looking to reduce inference latency without retraining from scratch may find its plug-and-play design useful, as the added linear attention branch and small LoRA adapters merge into the backbone weights at inference time, leaving the original model untouched.
diffuserssafetensorstext-to-videobase_model:MiniMaxAI/MiniMax-H3base_model:finetune:MiniMaxAI/MiniMax-H3 - 2Minimax H3_SingularityHF Models
WarmBloodAban· 128 likes
A fine-tuned fusion of MiniMax-H3 checkpoints, Minimax-h3\_Singularity supports text-to-video, image-to-video, video-to-video, and reference-to-video generation natively within ComfyUI. Builders working with multimodal video pipelines may find it worth evaluating given the reported weight pruning and optimization pass aimed at reducing artifacts introduced during high-step training.
minimax-h3video-generationtext-to-videoimage-to-videovideo-to-videoreference-to-video - 3stable-diffusion.cppGitHub
leejet· 6,930 stars
A C/C++ implementation of diffusion model inference covering Stable Diffusion, Flux, and several other architectures, built on top of the ggml tensor library. Builders targeting resource-constrained or CPU-first environments may find it useful as a dependency-light alternative to Python-based pipelines.
aicplusplusdiffusionfluxflux-devggml - 4H3 WorldHF Models
DANNY621· 71 likes
H3-World is a LoRA adapter (65.6M parameters, 0.199% of the 33B backbone) that sits on top of MiniMax-H3 to generate action-controlled video from a single starting frame, translating keyboard inputs into per-latent language instructions via directed attention routing. Builders exploring lightweight adaptation of large video diffusion models for interactive or game-world simulation will find it a concrete reference point for how little parameter budget such conditioning can require.
videoimage-to-videoworld-modelinteractive-world-modelloradiffusion - 5
promptslab· 6,317 stars
A hand-curated collection of prompt engineering resources focused on GPT-based models, ChatGPT, and PaLM, covering techniques like few-shot learning and prompt-based learning. Builders working on text-to-image or multimodal pipelines may find it useful as a reference for structuring and refining prompts before passing them to generation models.
chatgptchatgpt-apideep-learningfew-shot-learninggptgpt-3