MuseVRun

MuseVRun is the ComfyUI wrapper around Tencent's MuseV, the model behind half the "AI influencer" clips you saw in spring 2024. You feed it one portrait and it generates a short clip of that person moving - subtle head motion, eye blinks, a slight turn. If you want a talking-head avatar or a virtual-human asset, this is the node. If you want general video generation, don't come here: Wan and Hunyuan long ago outclassed it on arbitrary scenes. MuseV's niche is specifically virtual humans on a small SD 1.5 footprint, and in that niche it still has no real rival.

What it is

MuseV is a human-specific image-to-video model released by Tencent's TMElyralab team in March 2024. It rides the SD 1.5 ecosystem: a motion module plus a reference network (musev_referencenet) and IP-Adapter give the frozen base model temporal awareness and appearance lock-in. The pack's author, chaojie, did the minimal thing - the README is a paragraph and two workflow files, and the real documentation lives in the upstream TMElyralab/MuseV repo. Don't let the thin README worry you; the node works.

The output is intentionally small-motion: eye blinks, head waves, breathing. The default prompt ((eye blinks:1.8), (head wave:1.3)) tells you exactly what the authors trained it for. Fixed-camera character shots, not camera moves.

How it works

The node saves your ref_image to a temp file, then runs a two-phase pipeline. First a first-frame redraw: an img2img pass (first_redraw_steps, default 30) cleans up your input and establishes the look. Then the actual video pass runs parallel denoising - the motion module generates all frames together rather than frame-by-frame, which is why it avoids the error accumulation that plagued early AnimateDiff workflows. Appearance comes from the reference net + IP-Adapter keyed on your image, motion comes from the prompt and the training.

One honest gotcha from the source: video_len maps to time_size, the number of frames generated per batch (default 12, and the shipped workflow cranks it to 120). Each generated frame is also capped in motion - longer and wider means smaller motion, as the config comments note.

Inputs that matter

• ref_image (IMAGE) - your portrait. Feed it straight from LoadImage.
• prompt (STRING) - drives motion and style. The default is a good starting point for human subjects.
• video_len (INT) - total frames of the clip. 12 is short; 60–120 is typical.
• first_redraw_steps / first_redraw_guidance_scale - how much the first frame is redrawn (30 / 7.5).
• video_steps / video_guidance_scale - the video pass (10 / 3.5). Higher guidance = more motion, but more chance of the clip falling apart.
• sd_model_name - majicmixRealv6Fp16 (default) or fantasticmix_v10, both SD 1.5 checkpoints that ship inside the MuseV model download.
• uselcm (BOOLEAN) - flips on an LCM-LoRA for much faster generation at lower quality. Good for previews.

The output is a single IMAGE tensor - a batch of frames (T, H, W, C) - which you wire into VHS_VideoCombine to save as an mp4. That's exactly how the shipped workflow is wired. Skip the img_edge_ratio slider; the source hardcodes it to 1.0 in this version, so it's inert.

Install

Grab it from ComfyUI Manager (search "ComfyUI-MuseV"), or manually:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/chaojie/ComfyUI-MuseV
cd ComfyUI-MuseV
pip install -r requirements.txt

The --recursive matters: this pack vendors three submodules (MMCM, controlnet_aux, a diffusers fork), and a plain clone gets you an empty MMCM/. Then the big one - the model, several gigabytes:

huggingface-cli download --resume-download TMElyralab/MuseV --local-dir ComfyUI/models/diffusers/TMElyralab/MuseV

That one download contains everything: the motion UNet, the SD 1.5 base checkpoints, IP-Adapter weights, the LCM LoRA, and the negative embeddings. The pack's requirements.txt is short (decord, moviepy, h5py, …), but in practice you also need the pose/OpenMMLab stack (openmim, mmengine, mmcv==2.0.1, mmdet==3.1.0, mmpose==1.1.0) or the node refuses to import. Budget for a beefy GPU too - the community reported MuseV eating a full 24 GB back in 2024.

Common issues

• "Cannot be loaded" / import errors - almost always the missing MM packages or a clone without submodules. See install above.
• Model not found - the diffusers checkpoint must be at exactly ComfyUI/models/diffusers/TMElyralab/MuseV.
• OOM - drop width/height or video_len. 563×714 at 12 frames is a gentle start.
• Slow - enable uselcm and drop video_steps to 4–6 for iteration, then turn it off for final output.

It's a 2024-era model and it shows - but for one-photo avatar animation it remains the fastest path to something that looks alive.