Hunyuan Video Image To Guider

You've got one still image and you want a slow camera pan or a gentle zoom-in on it, as a video. And you're running Hunyuan Video, which at launch shipped text-to-video only - the official I2V model didn't land until three months later, and Wan had already eaten the ecosystem by then. This node is the stopgap: it takes your image and stretches it into a sequence of motion frames that "guide" Hunyuan's T2V weights into animating it. It's how a bunch of early Hunyuan workflows faked image-to-video before the real thing existed.

Don't overthink the name. The node doesn't output a GUIDER type at all - its single output is a batched IMAGE, and the "guider" part happens downstream when those frames get VAE-encoded and fed into a BasicGuider as conditioning. The name describes the job, not the wiring.

How it works

Read the source and it's ~200 lines of plain torch tensor math - no model weights, no API calls, nothing to download. For each of your frame_num frames it:

• Pans by converting the normalized move_range_x / move_range_y (−1.0 to 1.0) into pixel offsets and stepping across the image.
• Tiles seamlessly. Instead of running off the edge into black, it flips the image horizontally and/or vertically with torch.flip and wraps around, so a sideways pan tiles forever with no visible seam. This is the clever bit, and honestly the reason to use this node over a plain "crop and shift" utility.
• Zooms by center-cropping a shrinking window and bilinear-upscaling it back to full size, ramping the crop from 0 to your zoom value across the frames.
• Stacks the result into one tensor of shape (frame_num, H, W, C) that you hand to a VAE encoder.

The README's "Technical Details" section claims automatic mixed precision, batched processing, and progress tracking. The code does none of that. It's just interpolation, flips, and modular arithmetic - which is fine, because it's fast and it doesn't need the extra machinery.

The inputs that actually matter

All inputs are required, but you'll touch four of them:

• image - your still. Everything else is about how it moves.
• move_range_x / move_range_y - pan amount. The author's guidance is the one worth trusting: best range is −0.05 to 0.05. Crank it toward ±1.0 and each frame shifts by a full frame width; you get a strobe, not a camera move.
• zoom - 0.0 to 0.5. Keep it modest; this is a zoom in by default (ramps up over the clip).
• frame_num - 2 to 150. Match it to your sampler's frame count. The bundled example uses 69.

The rest you can leave alone until you need them: resize_mode (disabled / custom / keep_ratio) plus target_width / target_height to match Hunyuan's native resolution, and center_crop if you want a square crop instead of a resized one.

Wiring it up

Install via ComfyUI Manager (search "ComfyUI-HunyuanVideoImagesGuider") or the usual clone:

cd ComfyUI/custom_nodes
git clone https://github.com/benjiyaya/ComfyUI-HunyuanVideoImagesGuider

Then restart ComfyUI. There's no requirements.txt and no extra dependency - it uses only ComfyUI's bundled torch. The hard requirement is Hunyuan Video itself, and that's the real VRAM story: the 13B launch model wanted 45–60GB (FP8 helped), while HunyuanVideo 1.5 runs at 14GB with offloading. This node is the cheap part of the workflow.

Wire it like the included example: LoadImage → Hunyuan Video Image To Guider → VAEEncodeTiled → BasicGuider → SamplerCustomAdvanced, with a Hunyuan T2V checkpoint. The denoise setting is where people get burned: the README is explicit that 0.7–0.9 works and below 0.5 you get almost no object animation - the frames act as the start of the denoise, so too-low denoise keeps the image pinned and the motion dies.

The honest take

This is a thin utility node from a tiny pack (barely a whisper of community footprint - the author credits "Benji and DeepSeek AI Copilot"), and it shows: the code works, the README oversells, and there's no support infrastructure. It existed to fill a gap that the official Hunyuan I2V, then Wan, then Hunyuan 1.5's I2V all closed. Today you'd reach for it mainly if you're on legacy Hunyuan T2V weights and want a pan/zoom without a model swap, or if you specifically want that seamless mirrored-tile motion, which is genuinely neat. Otherwise, a proper I2V model does this better in one node. But for what it is - a free, dependency-free way to give a still image a camera move inside Hunyuan - it does the job.