RH MagiHuman Generate

This is the node the whole pack exists for. You feed it a reference image of a person, type what they should say and how they should say it, and out comes a talking-head video with synchronized audio - no separate lip-sync pass, no second TTS model to babysit. It's the payoff end of daVinci-MagiHuman, the GAIR-NLP model the community got excited about in spring 2026 mostly because of one thing: character consistency. Feed it a face it's never seen, have her turn away and come back, and it's still her. That's the killer feature against LTX and friends, and this node is how you reach it from ComfyUI.

How it works

The interesting bit about MagiHuman is the architecture. It's a 15B single-stream transformer where text, video, and audio all pass through the same self-attention - there's no "generate video, then sync a voice on top." The audio and the picture are decided together, which is exactly why the lip sync is so solid. That's also why you can hand this node a driving audio track if you want the mouth to follow a specific clip, but you don't have to: prompt alone gets you speech.

The node keeps it simple on purpose. It auto-detects orientation from your input image - portrait gets 256×448, landscape gets 448×256 - and pulls steps and guidance from the pipeline the loader handed it (32 steps with CFG for base, 8 steps with no CFG for distill). Your job is just the prompt, the length, and the seed.

Inputs that matter

• prompt - the whole game. Multiline, and worth structuring in three parts per the pack's guide: a main body (150–200 words covering appearance, facial dynamics, vocal delivery, camera), then Dialogue: <character, language>: "line", then Background Sound: <description>.
• seconds - 1 to 20, default 4. Short clips stay clean; long ones invite morphing.
• seed - default 42, and it has control_after_generate, so you can roll seeds from the widget.
• image and audio - both optional. No image and the model works from the prompt alone; feed it audio and the mouth follows it.

The output is a single VIDEO - 25 fps, with the audio muxed in - so it plugs straight into SaveVideo or PreviewVideo. Nothing else needed.

Installing it

Same as any custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-daVinci-MagiHuman.git
cd ComfyUI-RH-daVinci-MagiHuman
pip install -r requirements.txt

Restart ComfyUI. ComfyUI Manager can find it by the pack title too. The real install pain isn't the node - it's the weights. You need one INT8 DiT (~14.25 GiB) plus the T5-Gemma text encoder, the Wan2.2 VAE, and Stable Audio Open. The loader node's FileNotFoundError tells you exactly what's missing and where to grab it; the README's "Model Download & Installation" section has the full hf download commands.

Where people get burned

• The first run is the load, not the generate. Model loading on this pack is slow and eats VRAM; the generate itself is the fast part. Don't judge the node on the first execution.
• The built-in SR pass is heavy. The pack's own recommendation table says 540p SR wants ~24 GB and 1080p wants ~48 GB. On a 16 GB card, keep sr_model off - community testing repeatedly found the SR path the thing that turns "workable" into "slideshow."
• You need ffmpeg on PATH. The node shells out to ffmpeg to merge video and audio into the final MP4. It's a normal dependency, but a clean Python env won't save you if the binary is missing.
• Prompt structure pays. The three-part format isn't decoration; the model was trained to expect it, and dialogue lines you quote get spoken.

It's not an API node and needs no key - the name's a bit of a lie in that it isn't calling anything. It's just the generator end of a genuinely impressive open talking-head model, now small enough to run where you actually have a GPU.