Built a timeline video editor for MiniMax H3, inside ComfyUI 🎬
Write shots instead of graphs, keep your cast consistent with @<name>, extend or bridge clips seamlessly, cut picture + sound, export. An agent can drive it too.
Early preview, feedback welcome:
https://t.co/cJnZd6eni6
https://t.co/a3SfpG6jry
Welcome to Comfy for Developers.
This is where we’ll share:
→ Use cases from the community
→ Product improvements for developers
→ Developer info for every launch
→ Posts from our engineering blog
→ Hackathons and dev events
→ Freebies and swag giveaways
For company news, follow @ComfyUI. Subscribe to our dev-focused blog!
The first 50 to quote tweet us and fill out the form below will get $50 in Comfy credits (you need a Comfy account to get this!)
@nirjharhubay Good call. Swapping takes already keeps the slot and the cut where they are, so a longer take just fills the same space. What’s missing is showing how much of the new take is left over and letting you slip it inside the slot. Adding both.
New in v0.2.0, out today:
- LoRAs per project, with a strength slider each
- any H3 checkpoint, with its own steps, sampler and scheduler
- a different model per clip
- an agent can drive all of it over HTTP
Docs: https://t.co/jsKz3stZz9
H3 Editor is out: a video editor for MiniMax H3, inside ComfyUI.
Write shots, not graphs. The same cast across shots, a real timeline, fast drafts, large finals. Bring your own checkpoints and LoRAs.
Free and open source.
https://t.co/LYHkfXoruC
Every shot here was written as a sentence and rendered in the editor, sound included.
The film around them is code: built by Claude Fable 5.2 with our motion-graphics skill, also open source.
https://t.co/GEV7Lc7qUL
@TheXSami Yep! Each render is its own take, so you can redo just that shot and the rest of the edit stays put. Would love it if you gave it a try and told me what breaks. Thanks for the follow 🙌
Also in this release: one context pool for KV and state, sized by --context-capacity; a unified FP8 KV codec; JSON output that accepts the whitespace models naturally write; and automatic migration of retired startup flags.
Apache-2.0.
NInfer Workstation 2026.11.07 is out: a stable release of our Windows-native fork of the NInfer inference engine for Blackwell GPUs.
New Supervisor: settings on one searchable page, and a context slider that shows what each GiB buys before you restart.
https://t.co/HAQX4Lda4w
Speed with Qwen3.8-27B NVFP4 and speculative decoding, one request at a time, tokens/s:
RTX PRO 6000: code ~270, JSON ~390, prose ~165
RTX 5090: code ~215, JSON ~240, prose ~160
Each GPU uses its own NVFP4 build. 8 requests at once on the PRO 6000: ~530-590 tok/s total.
@dfolloni O Mempalace vai cuidar da estrutura e das buscas pelo lado dos agentes, mas para processar os docs talvez precise de uma pipeline tipo PaddleOCR ou similar, dá pra fazer tudo com modelos locais 👍
Israel, partial count (31% of the electorate, 3,906 ballot boxes, read 4:51 am Israel time): @PollsterGraph projection gives Likud 27 seats (80% range 25–28) and Netanyahu's bloc 56 (54–58); 61 or more in <1% of simulations. https://t.co/bU2jwgag7n