Prism Native Pipeline

What it is

Prism isn't a checkpoint. It's seven separate weight files - two video transformers, an audio transformer, a cross-modal bridge, a text encoder and two VAEs - that all have to be present, correctly matched, and loaded together before anything can generate. Prism Native Pipeline is where you assemble them into one object you can wire into a sampler.

You'd reach for it because it's the mandatory middle step: both Prism samplers take a single PRISM_PIPELINE input, and this is the only node that produces one. Skip it and nothing connects.

It's a metadata validator, not a loader (which is good news)

The most useful thing to understand: wiring this node does not pull 38 GiB into VRAM. The sockets carry component descriptors, and the weights only load at sampling time. Assembling the pipeline reads safetensors headers, checks the embedded component config, and tells you if you've got it wrong. That's why you can build the graph, hit the validation, and fix it in seconds instead of after a slow model load.

There are exactly three ways it rejects you, and they're worth knowing by heart:

• Something's missing. Missing Prism components: [...] - you left a socket empty. Note that the low-noise video DiT is optional in the runtime's own bundle check but still shows up as a required input on this node, so just wire all seven.
• A file is in the wrong socket. Incorrect component in <socket>: <kind> - you dropped the audio DiT into the video slot.
• The files aren't from the same conversion. Components are from different Prism variants/conversion bundles - every standalone component carries a bundle id, and all seven have to match. This is the one that catches people who mixed an alpha set with a beta set, or re-converted one file later.

If a file was never converted at all you won't even get this far: the loader nodes raise not a Prism standalone component. Convert official weights first. Sourcing the seven files from the same Hugging Face bundle is the simple way to stay out of trouble.

Inputs and output

Seven required inputs, no optional ones, no settings: video_dit, video_dit_2, audio_dit, dual_tower_bridge, text_encoder, video_vae, audio_vae. Each is its own socket type, so ComfyUI won't even let you mis-wire two loaders by accident - a Prism Audio DiT output only fits the audio input.

One output: PRISM_PIPELINE. It goes into pipeline on either Prism Native Video Audio Sampler (the full 50-step path) or Prism FreeVideo Accelerated Sampler (the 8/20-step distilled path). One pipeline node can feed both in the same graph if you want to compare them on one canvas - which is exactly the kind of A/B the pack's own example workflows are built for.

Install

ComfyUI Manager → search Comfyui-Prism-T8, or registry t8star/prism-t8, or:

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   # run with ComfyUI's own Python

Restart, then place the seven files: the four transformer components (prism_alpha_video_dit_int8_convrot, prism_alpha_video_dit_2_int8_convrot, prism_alpha_audio_dit_int8_convrot, prism_alpha_dual_tower_bridge_int8_convrot) in models/diffusion_models/, prism_alpha_text_encoder_int8_convrot in models/text_encoders/, and the two BF16 VAEs in models/vae/. The pack also registers its own models/ subfolders if you prefer to keep its files separate; models/standalone/ is the old flat layout and still works as a fallback.

Where people get burned

You mixed bundles. Nine times out of ten, "different Prism variants" means a half-updated folder. Re-download the full seven-file set rather than patching one file.

You pointed a loader at a Diffusers checkpoint. These are the pack's converted standalone files, not a Diffusers model directory. Configuration and tokenizer data are baked into each safetensors' metadata, which is the whole trick that makes Prism composable in ComfyUI at all. If you want to convert the originals yourself, the pack ships scripts/download_models.py and scripts/convert_models.py - budget around 72 GiB of source weights plus scratch space, and pick a fresh output directory per variant.

Nothing happens when you hit Run. The pipeline node is cheap; the sampler is where the CUDA requirement bites. Prism inference needs a CUDA GPU with BF16 support. On CPU it will load and then refuse, and the error is explicit about it.