RunningHub HeliosV2V

Video-to-video is where this pack earns its "transform existing videos with text prompts" line: feed it a clip, give it a sentence like "make it a rainy night, neon reflections," and get back a re-rendered version. What it is not is a quick filter. This is a full 14B diffusion pass over your footage - every frame gets regenerated with your prompt steering the style - so budget per-frame, not per-effect.

How it works

Mechanically it's HeliosT2V with a video input swapped in for image. The node pulls frames out of the input VIDEO object, feeds them to the pipeline as the conditioning video, keeps the first frame, and generates a new clip up to num_frames long under the guidance of your prompt. Same chunked generation, same 33-frame window, same optional pyramid stage2 acceleration - the only new piece is that your source frames are the backbone the model re-renders around.

One thing to get right up front: the video input expects a ComfyUI VIDEO type, not a batch of images. It reads .components.images off the object, so you need it to come from a loader that emits the VIDEO type - not a standard Load Video that hands you an IMAGE batch. If you're coming from the image world and only have image batches, you'll need to go through a video loader that produces the right type (or convert). This trips up more people than the actual generation does.

The inputs that matter

• video - the input clip. Its frames become the source of truth for structure and motion.
• prompt - this is the whole game. "Same scene, but at night" or "convert to oil painting" are the natural use cases; big prompt/source mismatches just produce mush.
• width / height - the output canvas. Your source gets used as-is as frames; keep these near the source resolution to avoid re-upscaling artifacts.
• num_frames - output length, multiples of 33 for clean chunks. Note it doesn't have to match the source length; the model re-renders whatever you ask for.
• guidance_scale - keep at the 1.0 default (x0-prediction distilled model, same as T2V/I2V).
• seed - your reproducibility friend once you find a look you like.
• is_enable_stage2 / pyramid_steps / is_amplify_first_chunk - the speed controls, unchanged.

Install

The shared pack install, once:

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

Model at ComfyUI/models/Helios-Distilled/, full details and the install.sh warning (it uninstalls xformers/torchao and pins triton) in the ModelLoader article.

Where people get burned

The VIDEO-type input is the number-one wall, so check that first. Second, temper the "transform" promise: with a short, low-frame source this node behaves a lot like I2V with extra context, and a long source at high resolution on an offloaded consumer card is a long wait. And third, this is regeneration, not preservation - expect the model's interpretation of your scene, not a pixel-faithful restyle. If you need the original motion preserved exactly, this isn't the tool; if you want a Helios-flavored reshoot, it's exactly right.