Prism High Noise Video DiT (UNet)

What it is

Prism isn't one file, it's seven, and this node is the one that loads the biggest of them: the high-noise video denoiser. Display name aside - there's no UNet in here, it's a diffusion transformer, the "(UNet)" is ComfyUI habit - this is the component responsible for the early, structural part of the denoising run: composition, motion, where things are.

You reach for it because you can't sample without it. It feeds the video_dit socket on Prism Native Pipeline, which feeds one of the two samplers. Everything downstream is downstream of this file.

High noise vs low noise, in one paragraph

Prism splits video denoising between two experts, the same shape of idea Wan 2.2 uses with its high-noise and low-noise models: one transformer handles the noisy early steps where the layout gets decided, the other cleans up detail once the structure exists. This node is the high-noise half. Its sibling, Prism Low Noise Video DiT (UNet), is the other. The handoff point between them isn't a setting anywhere - it's a boundary_ratio value stored in the bridge component's metadata, which the runtime reads when it wires the two together. So if you swap in a different bridge from a different bundle, you're also (correctly) moving the switch point.

The one input, and the one output

model_name is an enum listing whatever your models/diffusion_models/ folders contain. The file you want is prism_alpha_video_dit_int8_convrot.safetensors. Nothing else on the node - no dtype selector, no quantization toggle, no offload option. That's not an oversight: the sampler owns those decisions (int8_backend picks portable W8A16 or kitchen W8A8 for the video towers at sampling time), and the component's own metadata declares what it is.

What happens under the hood: the loader reads the file, confirms it carries a valid prism.format_version and that its declared component kind is video_dit, and passes a descriptor downstream. Actual weights stay on disk until a sampler runs - the sockets carry descriptors, not tensors, so building a graph doesn't cost you 38 GiB of VRAM.

ConvRot quantisation applies to the block Linear layers only. Embeddings, norms, time projections and the output head stay in floating point, which is why an INT8 file still loads at BF16 and still behaves like the model rather than a pile of rounding errors.

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   # ComfyUI's Python

Or ComfyUI Manager → Comfyui-Prism-T8 / registry t8star/prism-t8. Restart afterwards.

Download prism_alpha_video_dit_int8_convrot.safetensors from t8star/Prism-Comfy into ComfyUI/models/diffusion_models/. The full bundle is about 38.13 GiB across seven files; the two video transformers are the bulk of it. If you'd rather keep the pack's files out of your main model folders, Comfyui-Prism-T8/models/diffusion_models/ also works, and gets registered automatically. The old flat models/standalone/ still loads, but it's the legacy layout - prefer the categorised one.

Where people get burned

The dropdown won't show your file. Either it's in the wrong folder or it's not a converted Prism component. Loader errors are specific: not a Prism standalone component. Convert official weights first. means you pointed it at raw Diffusers weights or the base MOVA-360p files instead of the pack's converted output.

You dropped the low-noise file here. The two video DiT files are easy to mix up by name (prism_alpha_video_dit_int8_convrot vs prism_alpha_video_dit_2_int8_convrot). Put the low-noise one in the high-noise slot and the pipeline check fails with Incorrect component in video_dit: video_dit_2. Cheap to fix, confusing to read once.

You changed the file and ComfyUI reused a cached result. Won't happen - the loaders report file size and mtime as their change signature, so swapping files invalidates the cache. What does bite: mixing files from different bundles, which the pipeline rejects with a bundle-id mismatch. Keep all seven from the same download set.