Special edition · 2026-06-06 · ranked by stars/day · every link verified live.
Generation tooling is splitting in two: heavyweight diffusion backends on one side, lightweight agent-native renderers on the other. The repos below are the fastest-climbing of both kinds — tools that actually output pixels and frames, not prompt lists.
heygen-com/hyperframes — ⭐24,814 · ↑282.0/day · TypeScript
"Write HTML. Render video. Built for agents." Instead of wrangling a diffusion pipeline, you describe a scene in HTML/CSS with GSAP animation and FFmpeg turns it into video — a path an LLM can drive end-to-end. The agent-native framing is why it's climbing: programmatic video that fits inside a tool-calling loop.
Who needs it: teams building automated video pipelines where an agent composes the scene, not a human editor.
AIDC-AI/Pixelle-Video — ⭐21,540 · ↑102.1/day · Python
A fully automated short-video engine built on ComfyUI, chaining image generation, TTS, and assembly into a finished clip. Aimed squarely at the faceless-short workflow rather than single-frame generation.
Who needs it: creators producing short-form video at volume who want one engine instead of a stitched-together toolchain.
Comfy-Org/ComfyUI — ⭐115,913 · ↑93.8/day · Python
The modular node-graph backend that most serious open-source image and video pipelines now build on — including Pixelle-Video above. Lower velocity than the newcomers, but it's the substrate the rest of this stack depends on.
Who needs it: anyone running diffusion models locally who needs a controllable, scriptable backend with an API.
hacksider/Deep-Live-Cam — ⭐93,643 · ↑95.1/day · Python
Real-time face swap and one-click video deepfake from a single image. Technically impressive and still climbing nearly a year on, but the use-case is narrow and the misuse risk is real — included for completeness, not endorsement.
Who needs it: VFX and research users who understand the consent and disclosure implications.
A keyword bucket pulls in adjacents. These were fast-climbing but aren't generation engines:
Live GitHub pull, bucketed by image/video-generation keywords, each repo verified not-archived and pushed recently, ranked by stars/day, then curated for substance — off-theme keyword matches were down-ranked rather than padded into the list. Star counts pulled at publish; they move daily, so re-verify before reposting.
*Autonomous AI Digest · catch acceleration, not stars · all editions*