RunningHub HeliosI2V

If you're new to this pack, start here, not with T2V. Image-to-video hands Helios a real starting frame, which anchors the composition and makes the model's job easier - you get dramatically fewer of those "floating jellyfish that is not the subject you asked for" first frames that raw text-to-video produces. Give it a still you already like and a sentence about what should move, and you get a clip that actually looks like your image.

How it works

It's RunningHub HeliosT2V with one extra input. The image you feed in is the first frame; the pipeline generates the next num_frames − 1 from it, conditioned on your prompt, using the same chunked, pyramid-accelerated generation (33-frame chunks, rolling history, optional stage2 speedup). One honest gotcha in the code: the node takes your image and resizes it directly to width × height - no cropping, no padding. Feed a 1:1 square into a 640×384 canvas and it gets stretched, and the first frame looks off. Crop or resize your input upstream to match the output aspect ratio before it hits this node.

The inputs that matter

• image - an IMAGE tensor from anything upstream (Load Image, an SDXL/Wan still, a frame grabbed out of a video). Whatever is first in the batch gets used.
• prompt - now it's describing motion more than content: "waves rolling in, camera slowly pushing in" beats re-describing the subject.
• width / height - match these to your input image's aspect ratio (step 16). 640×384 default is fast but only right if your image is 16:9-ish.
• num_frames - multiples of 33 again: 99 ≈ 4s at 24fps, 132 ≈ 5.5s.
• guidance_scale - keep it near the 1.0 default; this is the same x0-prediction distilled model as T2V, and raising CFG mostly just blows out the still.
• seed - fix it once you find a motion you like.
• is_enable_stage2 / pyramid_steps / is_amplify_first_chunk - the pyramid speed controls, same as T2V, same tuned defaults.

Install

Same pack, same model, one time. Install via ComfyUI Manager ("ComfyUI_RH_Helios") or:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_Helios.git
cd ComfyUI_RH_Helios && pip install -r requirements.txt

The model goes in ComfyUI/models/Helios-Distilled/ (the full download + dependency warnings are in the ModelLoader article - including that install.sh uninstalls xformers/torchao and pins triton, so prefer the manual pip install in a venv).

Where people get burned

The stretch, already covered - that's the one that confuses everyone. Second: this is still a 14B model. On a 24GB card with offloading you're looking at minutes per clip, not seconds, so make your still count. Third, don't expect the model to invent motion the image can't support: if your input is a sharp, static product shot, Helios will animate it convincingly, but a heavily stylized or oddly cropped image drags the whole clip down. Feed it a clean, well-lit frame near the target resolution and the rest is prompt craft.

Output is the same video (VIDEO) object as the other nodes - an mp4 at 24fps written to ComfyUI's output folder - so wire it into whatever Save Video node you already use.