PAN 2 quantized from 81GB to fit a 16GB RTX 5060 Ti via custom ComfyUI nodes
Speedyrulz · reddit · 2026-11-12
A redditor ported IFM's 29B video model PAN 2 into a ComfyUI fork and quantized it for consumer cards. Additions include transformer model code with auto-detection, a Qwen3.5-9B text encoder type, Wan 2.2 VAE support, and new nodes (Text Encode, Negative Prompt, Image-to-Video, Spectrum, Mem Eff Sage Attention), plus a dual-GPU crash fix. Quantizations: int8 ConvRot at 39.8/27.4 GiB (0.9% weight error) and NVFP4 at 15.5 GiB (9% error), with CPU-only conversion scripts included. On an RTX 5060 Ti 16GB, the Spectrum node (running only 15/50 steps) cuts a 640x352/49-frame render from 149s to 58s; Sage Attention nearly halves step time at high resolution. Full-resolution 50-step renders take 10 minutes at CFG 2.5 without distillation. NVFP4 and CFG-1 runs show notable quality drops. Models/workflows are on Hugging Face; an experimental PAN 2 LoRA training branch is also available.
Related event: Tinkerer Shrinks 81GB PAN 2 Video Model to Fit 16GB GPU via ComfyUI(2 posts)→
More from Infra
- Agent harnesses are heading toward vertical integration, argues 0xblacklight — HankYeomans · 2026-11-12
- CoreWeave turns to customer prepayments and shorter contracts to fund AI growth — TechCrunch AI · 2026-11-12
- Foreign Investors Pull $50B From Asian Equities in October on AI Cash-Burn and Regulation Fears — nordicinst · 2026-11-12
- Bernstein Ranks Data Center Pipelines: META, MSFT, GOOGL Top the Credibility List — aronchick · 2026-11-12
- Photoroom reclaims ~1,000 idle GPU hours weekly using SkyPilot across Slurm and Kubernetes — skypilot_org · 2026-11-12
- UK AI chip startup Fractile raises $720M to challenge in the AI silicon race — TechCrunch AI · 2026-11-12