Prism FreeVideo Distill LoRA Pair

What it is

Two dropdowns and one output - that's the whole node. It picks up your high-noise LoRA, picks up your low-noise LoRA, and hands both paths to the accelerated sampler as a single PRISM_DISTILL_LORAS bundle.

The reason it's a pair node and not the usual "LoRA loader with a strength slider" is the way these particular weights are built. Prism carries two video denoisers - the video_dit expert and the video_dit_2 expert - and step distillation has to cover both halves of the denoising schedule or the early motion gets fast and the fine detail stays slow. So you get one file per expert. There is no strength widget here, and that's deliberate: the FreeVideo recipes expect the full pair at upstream values.

The two files

Both come from LightX2V's Wan 2.2 I2V line - the 260412 build, rank 256:

• Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_260412_rank_256_fp16.safetensors
• Wan_2_2_I2V_A14B_LOW_lightx2v_4step_lora_260412_rank_256_fp16.safetensors

Yes, they're Wan 2.2 LoRAs running against a Prism tower. That's not a mix-up: the FreeVideo engine applies this exact pair by name, and the pack's vendored notices cover Prism, MOVA, Wan and DAC together - the acceleration recipe documents the pair as required, so treat it as sanctioned rather than clever.

If you know how the community normally handles speed LoRAs on Wan 2.2, note the contrast: standard practice is to keep the high-noise pass clean and put the distill LoRA on the low-noise model only, at reduced strength. Here both experts get one, because the FreeVideo recipe is built around the paired distill.

Pick the pair up the easy way

python scripts/download_acceleration.py

That fetches both safetensors from Kijai/WanVideo_comfy at a pinned revision, verifies each one against its published SHA256, and drops them into the pack's own models/loras/ folder, which is registered automatically. It writes a lightx2v-source.json next to them recording the repo, revision and hashes, so you can tell later what you're actually running.

Manual route, if you prefer to keep things in the standard ComfyUI folder: pull the two files from Kijai/WanVideo_comfy → LoRAs/Wan22_Lightx2v at revision 8260d429 and drop them in ComfyUI/models/loras/. Also add the pack's models/loras/ to your mental map - the loader lists whatever folder_paths knows about, so either location shows up in the dropdowns.

Expect roughly 4.65 GiB for the pair. That's on top of the seven Prism components (~38 GiB) and, if you use the accelerated sampler, the 36.30 GiB streaming cache it prepares on first run.

Wiring it

Both high_noise_lora and low_noise_lora enumerate your loras folder. Put the HIGH file in the high slot and the LOW file in the low slot - the node won't stop you from crossing them, and nothing downstream will tell you it went badly. Then connect the output to distill_loras on the FreeVideo Accelerated Sampler.

Which tier needs it: Light, Standard and High all hard-error without the pair (Connect Prism Distill LoRA Pair: 260412 HIGH and LOW rank-256). Max does not use it at all - it's the undistilled 20-step path with joint CFG, so you can leave the node out of the graph entirely.

The native sampler ignores this node. If you want the original 50-step pipeline, you don't need any LoRA files, just the seven components.