HunyuanImageToVideo

Hunyuan Video was the first open video model that made local generation feel real - then Wan arrived, took the ecosystem, and by mid-2026 the standing answer to "should I use Hunyuan?" is a shrug. But the model still exists, it still has fans for its photorealistic lean, and this is the node that builds its image-to-video latent and conditioning. Notably, it's the only one in this family that takes just positive conditioning - no negative input on the schema. If you load a Hunyuan I2V workflow, this is the node doing the first-frame work.

How it works

Like its Wan counterpart, it creates an empty latent with 4x temporal compression, encodes your optional start_image with the VAE, and gets it into the pipeline. What's different is the guidance_type combo - three ways of telling the model "this is the first frame," because different Hunyuan checkpoints expect different mechanics:

• v1 (concat) - the classic approach: the encoded first frame is attached to the conditioning as a concat_latent_image with a concat mask, and the sampler sees it as extra context.
• v2 (replace) - the newer style: the image latent is written directly into the front of the latent sequence with a noise mask locking it, and a guiding_frame_index tells the model where the anchor lives.
• custom - exposes the raw reference latent (ref_latent) for you to wire up however the specific checkpoint wants.

If your workflow came from a template, the right value is usually already set; if you're assembling by hand and the model ignores your start image, flipping guidance_type is the first thing to try. The canvas inputs (width/height/length, defaults 848x480 at 53 frames, length stepping by 4) and batch_size behave like the rest of the family. Outputs are positive and latent, which go into a sampler with the Hunyuan checkpoint.

How you get it

Core node, no install. Models: the Hunyuan Video checkpoint (the 13B wants serious VRAM; the 1.5 line distilled it to 8.3B with a lower floor), the Hunyuan text encoder (a Llava-based CLIP), and for I2V the CLIP-Vision model - all in the standard models/ folders. The companion node TextEncodeHunyuanVideo_ImageToVideo is what produces the I2V conditioning, so check that one out if you're building from scratch.

The honest verdict

The community moved on for real reasons - Wan got the LoRA ecosystem, LTX got the speed, and Hunyuan's natural-looping party trick got closed by the VACE Video Joiner. Its niche that survived is realism enthusiasts who like the look and don't mind the license's territorial restrictions (the Tencent Hunyuan license excludes the EU, UK, and South Korea). If you're here because a workflow has this node in it, it'll run fine - just know you're on the legacy path, and the same effort on Wan 2.2 or LTX 2.3 will go further in 2026.