HUNYUAN VIDEO 1.5

Official Releases (from Tencent, pulled from Comfy-Org's HuggingFace)
Maintained and updated here for convenience.
Not affiliated with Tencent — just a fan of Hunyuan Video.

🎉 Training code for Hunyuan Video 1.5 is now released!
LoRA's incoming... Confirmed, Lora training working in musubi-tuner, its a bit picky, but it works.

I find that T2V works best with er_sde and beta11, and I2V works best with ipndm and simple/beta11.

TEXT TO VIDEO MODELS

• 720p T2V — FP16 — 16GB
• 480p T2V — FP16 — 16GB
• 480p T2V — FP8 — CFG-Distilled Scaled — 8GB

IMAGE TO VIDEO MODELS

• 720p I2V — FP16 — 16GB
• 480p I2V — FP16 — 16GB
• 480p I2V — FP16 — Step-Distilled — 16GB (added Dec 5)
• 480p I2V — FP8 — Step-Distilled Scaled — 8GB (added Dec 5)
• 480p I2V — FP8 — CFG-Distilled Scaled — 8GB

UPSCALE MODELS

• 1080p SR — FP16 — Distilled — 16GB

NOTES

• Text-to-Video models can also be used for Image-to-Video.
They behave differently from true I2V models, but still work.

TENCENT UPDATE LOG (summarized)

Dec 05, 2025
• 480p I2V Step-Distilled model released (8 or 12 steps recommended).
• End-to-end generation ~75% faster on RTX 4090 (≈75 seconds per video).
• Step-distilled quality remains close to the original.
• Optional 4-step mode available for ultra-fast output.
• Training code now open-sourced (Muon optimizer).
• HunyuanVideo-1.5 available on Hugging Face Diffusers.

Nov 27, 2025
• Cache inference support added (deepcache, teacache, taylorcache).
• Major speedups.

Nov 24, 2025
• Deepcache inference introduced.

Nov 20, 2025
• Inference code and model weights released.

LightX2V COMPATIBILITY

Tested and WORKING (4 or 8 step generation):
• T2V 720p FP16
• T2V 480p FP16
• I2V 720p FP16
• I2V 480p FP16
• I2V 480p FP8 CFG-D Scaled (Distilled)

Tested and NOT WORKING (full 50-step generation):
• T2V 480p FP8 CFG-D Scaled (Distilled)