HunyuanVideo 1.5 Leo Text-to-Video Sampler

HyVideo15T2VSampler is the text-to-video half of the simplified node set: load the model, type a prompt, get a clip. Same shape as its I2V sibling but no reference image - the model has to invent the whole scene from text. It's a two-node workflow with HyVideo15ModelLoader and it's the fastest way to test whether this pack (and the 1.5 model) is for you.

What it does

One node wrapping the whole pipeline - text encoder, transformer, scheduler, and the SR pass - plus the denoising loop. The settings you'll touch:

• hunyuanvideo_model_config - from HyVideo15ModelLoader. Set task to t2v there and match resolution (480p/720p).
• prompt / negative_prompt - write for motion and camera ("slow dolly-in", "camera pans right"), not just subject description. That's what separates a video prompt from an image prompt.
• video_length (121) - frames, ~7.5s at 16fps.
• num_inference_steps (50) / guidance_scale (6) - the base-model defaults. If you're on a distilled _distilled variant, drop steps to 8–12.
• aspect_ratio (16:9) - mapped to the nearest supported bucket.
• create_sr_pipeline (true) / sr_num_inference_steps (8) - the distilled 720p/1080p upscale pass. Real quality win; disable only for VRAM.
• prompt_rewrite + prompt_rewrite_base_url / prompt_rewrite_model_name - 1.5's prompt-rewrite feature, which sends your prompt to an OpenAI-compatible LLM endpoint for enrichment. The sampler only enables it when you give it a base_url and model name (it reads them via env vars), so with no endpoint configured, leave it off.

Same speed options as the I2V sampler: enable_cache (deepcache / teacache / taylorcache, needs angelslim==0.2.1) and enable_sageattn. Output is IMAGE (or raw latent if you set output_type to that).

The honest take

T2V is not where HunyuanVideo 1.5 wins. Community consensus across 2025–26 was consistently that 1.5's text-to-video underperforms its image-to-video, and that prompt adherence in T2V is the weak point - the distilled demo that got people excited was an I2V clip, and the model's reputation was built on I2V. If you primarily generate from text, Wan 2.2 or LTX are the stronger picks in the current landscape. If you're here anyway, keep expectations calibrated: give it a descriptive motion prompt and use the SR pass, because that's where 1.5's cinematic look shows up.

Installing it

ComfyUI Manager → "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 let HyVideo15ModelLoader auto-download (or point it at a manual model folder). That first download is huge - transformer, 7B text encoder, VAE, vision encoder, byT5.

Common issues

Same first-run pains as the I2V path: OOM (fix: enable_offloading in the loader, drop to 480p, disable SR), and the gated FLUX.1-Redux-dev vision encoder download failing (fix: HF token in the loader). One sampler-specific gotcha: negative_prompt defaults to empty here, so if your results come out noisy or unfocused, add the standard Hunyuan negative ("Aerial view, overexposed, low quality, deformation...") - the CFG node in the complete workflow ships with a good one you can copy.