ComfyUI ecosystem updates — open-source model releases, custom nodes, workflows, and tools for image, video, and audio generation.
Text to videoAlaya Lab releases EVOKE, a 14B open-weights world model generating 384x640 video with 3 steps, external world-state memory, and mid-flight prompt steering.
Apache-2.0
Text to videoByteDance releases Bernini-Diffusers-v2: the full semantic-planning video generation and editing pipeline in diffusers format, with stronger reference-guided editing and OpenS2V.
Apache-2.0
ComfyUI v0.33.1 brings native MiniMax Music 3 support with core CUDA Graphs, new MiniMax and Bria partner nodes, Anima tunes with extra blocks, and several stability fixes.
Image to videoLightx2v and ModelTC release MiniMax H3 Turbo Ref2V v0.1, a 4-step reference-to-video LoRA with ComfyUI-ready bf16 weights for multi-reference video generation.
Image to videoTencent ARC releases SCoPE, adding sightline-coordinate positional encoding to Wan2.2-I2V-A14B so videos follow a camera trajectory while keeping the image-to-video prior.
Apache-2.0
Text to videoLightricks releases LTX-2.5, a 22B open-weights audio-video foundation model with native multishot, 4K HDR, and official ComfyUI workflow templates.
Image to videoDownload MiniMax H3 Turbo LoRA v1.0: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors and the 4-step 768p ComfyUI variant, ready for models/loras.
Text to imageiljung1106's new ComfyUI node brings Normalized Attention Guidance to Krea 2, restoring negative-prompt control on Turbo checkpoints that require CFG 1.0.
Text to videoDownload h3-realism-people-t2v-i2v-r2v.safetensors (125 MB) into ComfyUI/models/loras for photoreal people in T2V, I2V and R2V with the r34l1sm trigger word.
Install the official MiniMax H3 skills from GitHub: h3-prompt-writing for all five generation modes plus eight style-specific video generation skills with bilingual SKILL.md files.
antirez, creator of Redis, builds h3.c, a native Metal-based MiniMax-H3 inference engine for Apple Silicon with prompt-to-video/audio and Ref2VA references working end to end.
MIT
Text to videoA ComfyUI custom node replaces MiniMax H3's 15.7 GB Qwen3-VL-32B text encoder with a Qwen3-VL-4B plus a learned projection, cutting conditioning VRAM to 4.5 GB.
MIT
InpaintingLanPaint 2.0.0 adds MiniMax H3 support with a video and audio inpainting pipeline: paint per-frame masks and audio intervals in one editor, then inpaint both in a single pass.
Text to imageLodestones releases Kroma v0.2 as a full Krea 2 fine-tuned checkpoint instead of a LoRA, with base and turbo files plus community INT8 quants for ComfyUI.
MIT
Video editingH3 Motion Context v0.2.0 chains MiniMax H3 clips so motion and audio continue across cuts, adds reference mode support, and removes the visible seam at joins.
Text to videoMamad8 releases a 2x clean-latent upscaler for MiniMax H3 with two ComfyUI nodes, doubling spatial resolution before VAE decode while staying in latent space.
Text to videoThe v4 step-600 Turbo LoRA distills MiniMax H3 audio-video generation to 2-3 sampling steps, with updated ComfyUI nodes supporting pruned checkpoints and a bundled workflow.
Image editingMamad8 releases an experimental image-specialized MiniMax H3 VAE that decodes a single T=1 temporal latent directly into an image, no custom node required.
Character animationWan-Animate-2 and Wan-Animate-2-Lite for ComfyUI: character animation from driving video, text-driven viewpoint control, native nodes, real-time Lite for avatars.
Apache-2.0
Image to videoLightx2v and ModelTC distill MiniMax H3 into a 4-step FL2V Turbo LoRA, cutting sampling from ~20 steps to 4 with a ComfyUI conversion and demo Spaces.