Prism Low Noise Video DiT (UNet)

What it is

The second of Prism's two video denoisers. Where the high-noise expert decides composition and motion, this one takes over late in the schedule and cleans up - texture, edges, faces, the parts that look mushy if you skip it. Two experts, one handoff, one output file each.

You reach for it because it's a required part of the stack: it feeds the video_dit_2 slot on Prism Native Pipeline. Two-expert video models are the norm now (Wan 2.2 split its denoising the same way, and that's why the community's speed-LoRA advice has always been "low-noise pass only, reduced strength"), and Prism follows it.

What's actually on the node

One input: model_name, an enum of files found in your models/diffusion_models/ directories. One output: video_dit_2. No settings, no dtype picker, no offload switch - all of that lives on the sampler, which decides dtype and quantisation backend for the whole video stack when it runs.

The file you want is prism_alpha_video_dit_2_int8_convrot.safetensors. Note the _2_ in the middle. The high-noise file is prism_alpha_video_dit_int8_convrot.safetensors, no digit, and putting them in each other's slots is the single most common way to fail the pipeline check.

Two things worth knowing about the mechanism:

The switch point isn't yours to set. The handoff between high- and low-noise experts is driven by a boundary_ratio value that lives in the bridge component's metadata. The runtime reads it when it assembles the pipeline. So the split is a property of the converted bundle, not a slider - if you want different behaviour you'd need a differently converted bridge, and then all seven files should come from that bundle.

This one is optional in the runtime, required on the node. The pack's internal bundle check treats video_dit_2 as the one component that may be absent, but Prism Native Pipeline still exposes it as a required input. Wire it and stop thinking about it. Every example workflow does.

Like its sibling, the ConvRot quantisation here covers the block Linear layers; embeddings, norms, time projections and the output head stay in float. The file also declares its own format version and component kind, which is what lets the pipeline reject mismatched files instead of producing garbage.

Install

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

Or ComfyUI Manager → Comfyui-Prism-T8, or registry t8star/prism-t8. Restart ComfyUI after installing.

Then grab prism_alpha_video_dit_2_int8_convrot.safetensors from t8star/Prism-Comfy into models/diffusion_models/. Download all seven files from that repo in one go - the bundle-id check means a partial or mismatched set is rejected at the pipeline node, and a lone re-download is the usual cause.

Where people get burned

Wrong slot. Incorrect component in video_dit_2: video_dit - you swapped the two video experts.

Different bundle. Components are from different Prism variants/conversion bundles - the seven files must all come from the same alpha or beta conversion. This shows up most often when someone converts one component themselves and mixes it with a downloaded set.

Nothing in the dropdown. The file isn't in a folder ComfyUI scans for diffusion models, or it isn't a converted component. The loader says so directly: not a Prism standalone component. Convert official weights first. Raw MOVA or Diffusers weights won't load here.

Expecting speed from the drop. Skipping the low-noise expert isn't a supported optimisation - its presence is assumed by the recipes, and the distilled LoRA pair that makes the accelerated sampler fast includes a low-noise half precisely because this tower still runs.