RunningHub HeliosT2V

This is the headline node of the pack: give it a sentence, get back a video. Wire a HELIOS_PIPE from RunningHub HeliosModelLoader into helios_pipe, type a prompt, connect the video output to a Save Video node, hit run. That's the whole workflow.

The "real-time" claim in every Helios headline - 19.5 FPS on a single H100 - is doing a lot of work. On a consumer card this node runs the same model far slower, and you should expect to wait minutes for a decent clip, not seconds. The community reaction to Helios landing was exactly that: awe at the demo, then "so when's it run on my 4090." This node is RunningHub's answer to that - it makes the model usable, not real-time. Set expectations and it's a lot of fun.

How it works

Helios doesn't generate all 480 possible frames in one pass. It generates in chunks of 33 frames - the window is (9 latent frames − 1) × 4 + 1 - and carries a rolling history of previous frames ([16, 2, 1] frames short/mid/long) into the next chunk so long clips stay temporally coherent. That's why the README pushes multiples of 33: 99 frames ≈ 4s at 24fps, 132 ≈ 5.5s, 264 ≈ 11s.

Two knobs live on top of that chunking. is_enable_stage2 turns on the "pyramid" acceleration: later chunks are denoised at progressively lower resolutions and upsampled, which is most of the speed win. pyramid_steps (default "2,2,2") sets how many denoising steps each of the three pyramid stages gets - the sum, 6, is what each non-first chunk effectively runs. is_amplify_first_chunk doubles the steps on the very first chunk, because the opening shot sets the whole clip's quality. On the distilled model these are tuned defaults; leave them alone until you know why you're changing them.

The inputs that matter

• prompt - multiline, and it does the work. Be specific about motion, camera, and subject.
• width / height - 640×384 by default, stepped in multiples of 16 (width 128–1920, height 128–1088). Stick near the defaults for speed.
• num_frames - default 99, max 480. Use multiples of 33; a non-multiple just pads a partial chunk.
• guidance_scale - default 1.0, and this is the distilled-model tell. Helios-Distilled is an x0-prediction model trained with a custom scheduler; you don't dial CFG up like an SD checkpoint. Crank it toward 2+ and you'll blow the image out. If you swap to Helios-Base (v-prediction, standard CFG), that's when you raise it.
• num_inference_steps - 50 for the first chunk (then pyramid steps take over). Fewer is faster, rougher.
• seed - set it if you want a reproducible shot.
• negative_prompt - optional, and on the distilled model it does surprisingly little. Don't build elaborate negative chains; the model mostly doesn't need them.

Install

Install the pack and model once (see the ModelLoader article for the full version, including the model download and the numpy<2.0 / triton caveats):

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

Then place Helios-Distilled at ComfyUI/models/Helios-Distilled/ - huggingface-cli download BestWishYsh/Helios-Distilled --local-dir ComfyUI/models/Helios-Distilled.

Where people get burned

The output is a ComfyUI VIDEO object - the node exports an mp4 to ComfyUI's output folder at 24fps and wraps it, so make sure your Save Video node expects that type. And don't judge the model on your first 99-frame run with offloading on a 8GB card: the progress bar is honest about total steps (it estimates them from the pyramid math), but each step is slow under group offload. Give it one solid multi-minute run before you start tweaking prompts - the first impression of Helios from a cold, offloaded start is its worst one.