PAN2 29B Video Model Running in ComfyUI: Quantized to 15.5GB With 2.5x Speedups
Speedyrulz · reddit · 2026-11-11
Impatient for official support, Reddit user Speedyrulz built a ComfyUI fork adding PAN 2 (IFM's 29B video model, 81GiB at release), with quantized checkpoints and new nodes, tested on an RTX 5060 Ti 16GB.
Added to ComfyUI
- Model code for the PAN 2 transformer (44 joint blocks + 4 text refiner blocks) with auto-detection
- New pan2 text encoder type (Qwen3.5-9B), Wan 2.2 VAE compatibility
- New nodes: Text Encode, structured Negative Prompt, Image To Video, Spectrum, Mem Eff Sage Attention
- Fix for dual-GPU positive/negative pass crashes
Quantizations (conversion scripts included, CPU-only)
- int8 ConvRot: 39.8 / 27.4 GiB (0.9% weight error)
- NVFP4: 15.5 GiB (9% error, native on RTX 50 series)
- Qwen3.5 encoder: 8.4 GiB int8 / 4.6 GiB NVFP4
Measured speedups (RTX 5060 Ti)
- Spectrum node (forecasts steps: only 15/50 actually run): 149s → 58s
- Mem Eff Sage Attention at full res: 42.2 → 25.7 s/step
Tips: 50 steps, CFG 2.5, euler/simple, 24fps recommended; PAN 2 expects structured JSON prompts — a plain-text negative prompt visibly degrades quality. A 10s 1MP sample took 26 minutes, then frame-interpolated to 48fps.
Related event: Tinkerer Shrinks 81GB PAN 2 Video Model to Fit 16GB GPU via ComfyUI(2 posts)→
More from Infra
- KubeCon keynote: computing is shifting from human speed to agent speed — HowDevelop · 2026-11-12
- Leaked memo: Meta infrastructure head Santosh Janardhan to leave in early 2027 — dinabass · 2026-11-12
- Huawei open-sources MACaQ quantizer: 23% lower KL than Unsloth GGUFs — acluk90 · 2026-11-12
- Omarchy 4.6 lets you share local AI models across your tailnet, says DHH — alexcovo_eth · 2026-11-12
- Eric Schmidt's foundation funds French NGO fighting against data centers — StephanSturges · 2026-11-12
- Epoch: Open-Weight GLM 5.3 Hacks Six Firefox Versions for ~$1,000 Each; Claude Fable 5.2 Sets ECI Record — Epoch AI · 2026-11-12