PAN 2 29B world model runs in Forge Neo with a timeline editor that writes its JSON prompts
digitalhunters0 · reddit · 2026-11-18
A developer built two Forge Neo extensions for PAN 2, the 29B video world model from IFM (MBZUAI):
- PAN 2 for Forge Neo: text-to-video and image-to-video (image as first frame) from the standard tabs, 24fps, up to 15 seconds, with the official recipe preset (Euler, 50 steps, CFG 2.5, Shift 9) and support for speedyrulz's INT8/NVFP4 files, Qwen3.5 text encoder, and Wan 2.2 VAE.
- Regional Boxes timeline editor: PAN 2 relies on second-by-second structured JSON descriptions (plain sentences work noticeably worse). The extension turns prompt-writing into a visual timeline — fill style/lighting/subjects, drag to split events, pick camera moves from PAN 2's own list, and get JSON generated for you, plus a pre-flight check for missing fields.
Honest caveats: it's slow — 17s/step at 1248x704 on an RTX 6000 Ada (48GB, INT8), so a 5s clip takes 15 minutes; 640x384 is 4-5 minutes for testing. No audio; no official prompt rewriter or GGUF yet. Model is Apache 2.0.
More from Multimodal
- Adobe Firefly ambassador shares full workflow for embroidery-style AI animated short — azed_ai · 2026-11-18
- PixVerse-generated Vice City sunset street-fight clip goes for GTA vibes — azed_ai · 2026-11-18
- Google open-sources StreetReaderAI, a multimodal map app for blind street exploration — jonfroehlich · 2026-11-18
- Rumble-2-Flash: open-weights speech-to-text model handles accents, noise and 40+ languages — omarsar0 · 2026-11-18
- Runway cuts real-time 57B video model serving from 8 H100s to 4 with inference optimizations — runwayml · 2026-11-18
- Open-weights ASR model Rumble-2-Flash hits #9 on Hugging Face Open ASR leaderboard, 40+ languages — cneuralnetwork · 2026-11-18