Hunyuan

Start with HunyuanVideo 13B via ComfyUI on RTX 4090 24 GB — HunyuanVideo is Tencent's open-weight text-to-video model generating 720p 24fps video clips up to 129 frames (~5 seconds) in ~15 minutes on a single RTX 4090. It uses a dual-stream DiT architecture with 3D VAE and CLIP + Llama 3 text encoders, producing higher motion coherence than Wan 2.1 at the cost of 2× generation time. For 720p, HunyuanVideo is the quality leader among open-weight video models — motion smoothness and prompt adherence exceed Wan 2.1 on the VBench benchmark. The 13B DiT model at FP16 is ~26 GB — must use FP8 attention (--fp8_e4m3fn) on 24 GB cards or upgrade to A6000 48 GB. For lower VRAM, there is no practical sub-13B option — HunyuanVideo requires 24 GB minimum. Apache 2.0 license.

For single-user generation: ComfyUI with HunyuanVideoWrapper node on RTX 4090 24 GB with FP8 attention — ~15 min for 129-frame 720p clip at 30 denoising steps. The 3D VAE spatial+temporal compression produces ~3 GB latent per clip — tile-based encoding (temporal_tile_size=32, spatial_tile_size=256) prevents OOM. For A6000 48 GB: full FP16 with batch size 1, ~12 min per clip. For multi-GPU: ComfyUI + model parallelism across 2× RTX 4090 with --device-map auto — ~10 min per clip. For server/production: deploy as batch job processor with GPU queue — video generation latency is incompatible with real-time serving. The dual text encoder (CLIP ViT-L/14 + Llama 3 8B) occupies ~8 GB combined at FP16 — keep both encoders in GPU memory; offloading to CPU adds 3-5 seconds per encode. For maximum quality: use 50 denoising steps + DPM-Solver++ scheduler — extends generation to ~25 minutes but yields best VBench scores.