Prism FreeVideo Accelerated Sampler

What this node is for

Tencent's Prism is one of the few open models that generates a video and its soundtrack in the same denoising pass, instead of generating silence and gluing MMAudio onto it afterwards. The catch is that the native path is a 50-step, 1280×720 beast, and on a normal GPU that's a coffee break measured in kettle refills.

Prism FreeVideo Accelerated Sampler is the fast lane. It hands your generation to the pinned FreeVideo engine - a step-distilled fork of Prism's own sampler - and runs it in a separate worker process with its own memory budget. You pick a quality tier, you get a recipe that has already been tuned for you. You will not get the full-quality 50-step result. You will get roughly the same scene in a fraction of the time, which is what you actually wanted on a Tuesday night.

What the quality tiers really change

This is the part people get wrong. quality is not a speed slider: it's the entire recipe switch, and the example workflows' node titles are just labels - the dropdown is the truth.

Light, Standard and High all run the same video pass: 8 steps, CFG 2 applied only on the first step, shift 5. The difference is where the audio comes from as the video denoises.

| Tier | Video | Audio guidance | | --- | --- | --- | | Light | 8 steps, CFG 2 on step 1 | The student's own v2a K/V, 4 audio sub-steps at CFG 5 - no extra teacher pass | | Standard | same | Partial teacher: fused layers 0–9 keep the student's K/V, layers 10–29 come from a base pass | | High | same | Full teacher over all 30 fused layers, every step | | Max | 20 steps, CFG 5 all steps | Joint video/audio CFG; no distillation LoRA involved |

High is the node's default. The vendored tier table inside the pack marks Light as its default, which tells you something about how much the author trusts the cheap audio guides.

The inputs you actually touch

Wire pipeline from a Prism Native Pipeline node - that's the whole model stack in one cable.

prompt and audio_prompt are separate on purpose: describe the picture in one, the sound in the other. If you leave audio_prompt empty it reuses the video prompt, which is how the upstream Mandarin comparisons are done - Chinese scene prompt, dialogue wrapped in <speech>…</speech>, audio prompt blank.

width, height, num_frames, fps, seed. The author's own tooltips are blunt about this: 1280×720 at 205 frames / 24 fps is the upstream reference, and smaller sizes are experimental and may degrade audio. Short clips too. This isn't a node that likes being pushed around.

vram_gib and ram_gib (18 / 20 by default) are placement budgets passed to the worker, not hard ceilings. If you have a 24 GB card and nothing else resident, raising vram_gib lets it keep more blocks on the GPU instead of streaming them.

reference_image is the one that trips people: Light, Standard and High require it. Only Max runs from text alone - disconnect the image there.

distill_loras takes the output of Prism FreeVideo Distill LoRA Pair. Light, Standard and High won't start without it; Max ignores it.

Outputs are frames (IMAGE), audio (AUDIO) and fps (FLOAT). Frames and audio go into Prism Save Video with Audio; that's the entire second half of the graph.

Install

ComfyUI Manager, search Comfyui-Prism-T8, or the registry entry t8star/prism-t8. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Prism-T8.git
cd Comfyui-Prism-T8
python -m pip install -r requirements.txt   # same Python that runs ComfyUI

Then download all seven standalone model files from t8star/Prism-Comfy - about 38.13 GiB - into models/diffusion_models/, models/text_encoders/ and models/vae/, and fetch the acceleration LoRAs:

python scripts/download_acceleration.py

Restart ComfyUI afterwards. You need a BF16-capable NVIDIA CUDA GPU, Triton matching your PyTorch (on Windows, triton-windows), and comfy-kitchen. The tested environment is Python 3.10, PyTorch 2.11/cu128, triton-windows 3.6.

Where people get burned

First run looks hung. The worker builds a private streaming cache in the plugin's models/.prism-acceleration-cache/ before sampling - 36.30 GiB with the alpha bundle and both LoRAs. It's prepared one unit at a time, on CPU, and reused afterwards. Change any input file and it builds a fresh one. Budget disk, and clear old cache dirs.

The speed numbers are somebody else's hardware. The 14.6× figure is upstream's, measured on an H200 against the 50-step original, excluding loading. The pack's own laptop reference: 768×432 / 121 frames / 8 steps took 311.8 s, and 1280×720 / 205 frames took 2,210 s on an RTX 5090 Laptop. Plan accordingly.

The audio can be wrong for reasons that aren't your fault. The user and the FreeVideo author jointly reproduced input-dependent audio degradation in Prism's base model for certain prompt/seed/reference combinations. Listen to the track. Decoding cleanly proves nothing.

Note also that there's no portable/kitchen choice here - the worker runs FreeVideo's own W8A8 diffusion recipe, and the portable/kitchen switch belongs to the native sampler.

One more thing worth knowing: this model has essentially no English-language trail. Searches for "FreeVideo" turn up nothing in the Reddit side of the community, and the pack's docs and demos live on the author's Bilibili channel. If something looks odd, you're mostly on your own with the README.