I made ComfyUI nodes for LongCat-Video-Avatar 1.6: one photo + one voice = lip-synced talking video (free, MIT)
retained text
the 3 clips are one run each: one photo + one ElevenLabs voice line, 480p, seed 42. no lora, no face swap, no editing. sound on. built it on top of rookiestar28's 1.5 pack (MIT, full credit to them). what's different in 1.6: the official INT8 model, 8 steps, the 1.6 LoRA switch (mix\_steps), no loudness norm on the audio, and it picks SageAttention > FlashAttention 2 > PyTorch by itself. long videos work window by window (93 frames, about 3.7 s each). the model files download on first run (about 25 GB), you only add the umt5 text encoder. honest numbers: I only tested on a 5090 32GB + 64GB RAM, Windows, 480p. a 3.8 s clip takes 5.6 to 7.7 min depending on attention and mix\_steps, each extra 3.2 s of audio adds one window (1.5 to 2.5 min). I haven't tried cards under 32GB. with block streaming the model itself used about 13GB at 480p, but that's not the same as "runs on 12GB". what helped: calm presenter = write "hands relaxed" in the prompt, clean speech with no music, short 4 s tests first, new seed instead of rewriting the prompt.
document fetched 2026-11-19 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xzjx31/i_made_comfyui_nodes_for_longcatvideoavatar_16