HunyuanVideo 1.5 Leo Image-to-Video Sampler

Image-to-video is where HunyuanVideo 1.5 genuinely shines, and HyVideo15I2VSampler is the one-click path to it. This is half of the pack's "simplified" set - the other half being HyVideo15ModelLoader - and together they're a two-node workflow: load the model, feed an image and a prompt, get a video. If you've come to this pack because someone showed you a 40-second 1.5 I2V clip, this is the node they used.

What it does

Everything, in one place. It builds the full pipeline internally - VAE, text encoder, vision encoder, transformer, scheduler, and the SR pass - then runs the denoising loop and hands back a finished IMAGE tensor. The inputs you actually set:

• hunyuanvideo_model_config - from HyVideo15ModelLoader. This is the whole loaded model stack; nothing works without it.
• reference_image - your starting frame. This is I2V, so no image means no video.
• prompt / negative_prompt - the motion and scene you want. Keep the prompt descriptive of motion ("the camera slowly pushes in...") since I2V already has the subject locked by the image.
• video_length (121) - frames. The model's default; at 16fps that's ~7.5 seconds.
• num_inference_steps (50) - denoising steps. The base model wants 50; the step-distilled 480p_i2v_distilled variants are built for 8–12 and are about 75% faster, which is the recommended fast lane.
• guidance_scale (6) - CFG scale, >1 enables classifier-free guidance.
• aspect_ratio (16:9) - the output bucket; the sampler maps it to the nearest supported resolution for the model version.
• create_sr_pipeline (true) - runs the distilled 720p/1080p super-resolution pass after the base generation. The quality jump is real, and sr_num_inference_steps (8) keeps it cheap. Disable it if VRAM is the wall.
• output_type (pt) - pt gives you frames for ComfyUI's video nodes; latent returns raw latents for downstream fiddling.

The optional extras are speed toggles: enable_cache with cache_type (deepcache / teacache / taylorcache) and enable_sageattn, which trade a little fidelity for big speedups via cached blocks and SageAttention. They need angelslim==0.2.1 (in the pack's requirements) installed, and enable_sageattn needs SageAttention itself.

Why you'd pick 1.5 for this

The honest context: in 2026 the local video ecosystem runs on Wan and LTX, and Hunyuan 1.5's T2V is its weak leg. But its I2V - especially the step-distilled build - is fast, prompt-faithful, and has a distinctly cinematic default look, and the 201-frame clean-loop trick it inherited is still a fun party piece. The community demo that got people talking was a 40-second I2V clip on a 5090. On more modest cards, enable offloading in the loader and drop to 480p.

Installing it

ComfyUI Manager → search "HunyuanVideo-1.5 nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yuanyuan-spec/comfyui_hunyuanvideo_1.5_plugin
cd comfyui_hunyuanvideo_1.5_plugin
pip install -r requirements.txt

Restart. Then in HyVideo15ModelLoader set model_path empty (or the folder path) and task to i2v, pick resolution, and run. First run downloads a lot: the 8.3B transformer, the 7B Qwen text encoder, the siglip vision encoder, the VAE, and byT5. The vision encoder comes from FLUX.1-Redux-dev, which is gated on HuggingFace - if the download errors on that piece, accept the license and paste your hf_token into the loader.

Common issues

Out-of-memory is the classic first-run wall - turn on enable_offloading in the model loader and consider disabling create_sr_pipeline. The other common miss: a reference image that's wildly off-aspect from your aspect_ratio setting, which gets center-cropped hard. And remember the license: HunyuanVideo 1.5 is restricted outside the EU, UK, and South Korea. Fine locally; not fine for a product you ship everywhere.