HunyuanVideo Prompt Mix Sampler

Every other sampler in this pack takes one hyvid_embeds input. This one takes two - hyvid_embeds and hyvid_embeds_2 - and blends between them during generation rather than making you pick one prompt or the other. That's a genuinely different tool from string-level prompt concatenation: mixing two prompts as text ("a cat, a dog") just gives the model a confused combined description, while mixing them in embedding space (what this node does) is closer to interpolating between two concepts the model actually understands separately.

Unlike the inversion sampler pair, this one does a normal from-scratch generation - no source video, no inversion step. It needs the standard txt2vid setup: width and height (both default 512), num_frames (default 17), steps (default 30), plus seed and force_offload like any sampler. Don't confuse this with the HyVideoInverseSampler/HyVideoReSampler pair - this is generating something new that transitions between two concepts, not editing an existing clip.

The one field unique to this node is alpha (0–1, default 0.5), and its tooltip is straightforward: it "adjusts the blending sharpness." Read that as the interpolation curve between your two prompts - 0.5 sits at an even middle, and pushing toward 0 or 1 leans the blend toward whichever hyvid_embeds input corresponds to that end. What you get in practice depends heavily on how different your two prompts actually are: two closely related descriptions blend smoothly into a coherent single subject, while two wildly different ones can produce something that reads as a transition or morph across the clip rather than a stable "third thing." Worth experimenting with both - sometimes the morph effect is exactly what you want.

embedded_guidance_scale defaults to 6 and flow_shift defaults to 9 here - notably higher than the 1 you'll see on the inversion samplers. That's a real difference, not an oversight: this node is generating fresh content rather than following a recovered trajectory, and the pack ships it with defaults tuned for that. Leave them where they are unless you have a specific reason to move them; flow-matching schedulers like this one are generally more sensitive to their shift value than beginners expect, and drifting it without understanding why tends to produce worse results than just leaving the shipped default alone.

Two optional inputs, both shared with the inversion samplers: interpolation_curve varies effect strength across the temporal axis (per-frame rather than per-step), and feta_args is the same optional Enhance-A-Video-style detail/adherence boost you'll see wired into HyVideoReSampler - skip it if you don't already have that node in your graph.

Installing the pack: ComfyUI Manager, search "HunyuanVideoWrapper," or by hand - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-HunyuanVideoWrapper, then pip install -r requirements.txt inside the cloned folder before restarting ComfyUI. Skipping the pip step is the single most common reason people end up with a partial node set after installing through Manager alone.

A realistic expectation to set going in: prompt mixing at the embedding level is a genuinely fun, experimental tool, but it's not going to give you frame-precise control over exactly when the transition happens - that's what alpha roughly steers, not something you dial in with surgical precision. Treat your first few runs as exploration, not a repeatable production technique, until you've got a feel for how your specific two prompts behave together.