Wan

Start with Wan 2.1 T2V 14B via ComfyUI on RTX 4090 24 GB — Wan 2.1 is the best open-weight text-to-video model as of mid-2026, generating 5-second 480p clips at 16fps in ~8 minutes. The 14B variant uses a DiT-based video diffusion architecture with 3D VAE that compresses spatial+temporal dimensions. For image-to-video, use Wan 2.1 I2V 14B — same architecture, different conditioning. For lower VRAM (<16 GB), use the 1.3B variant — generates 480p in ~2 minutes on RTX 3060 12GB but with visible quality degradation in motion coherence. Skip Wan 1.0 — the 2.1 architecture added tile-based VAE encoding that prevents OOM on consumer GPUs. Wan uses Apache 2.0 license — no commercial restrictions. For higher-quality video generation at datacenter scale, compare HunyuanVideo.

For single-user generation: ComfyUI with WanVideoWrapper node + Wan 2.1 T2V 14B FP16 on RTX 4090 24 GB. The 3D VAE encodes video frames into latent space at 16× spatial and 4× temporal compression — entire 5-second clip latent is ~2 GB. For 24 GB cards: use --fp8_e4m3fn attention + tile-based VAE encoding (tile_size=256). For 48 GB cards (A6000/L40S): full FP16 with batch size 1 works without tiling. For server/production: ComfyUI API mode with GPU queue — video generation is too slow for real-time serving; treat as batch job processing. The 14B DiT model at FP16 is ~28 GB — requires at minimum RTX 4090 with FP8 attention or 2× RTX 3090. For LoRA training: ~20 GB VRAM for rank-16 LoRA on 14B model — requires A6000 48 GB minimum. Generate at 480p and upscale with Real-ESRGAN post-process instead of generating native 720p (4× generation time for marginal quality gain).