Load Bernini v2 Planner

Every Bernini workflow you saw in 2026 ran the renderer and crossed its fingers. Kijai, who wrote the standard ComfyUI integration, said it plainly: he tried the planner and couldn't get anything out of it that the renderer-only model didn't already give him. So the community ran Bernini-R - the Wan 2.2 expert fine-tune - and the Qwen-based semantic planner that's half the paper stayed theoretical. Load Bernini v2 Planner is the node that closes that gap: it loads the actual Qwen2.5-VL language/vision modules, the Bernini connector, and the VIT flow decoder, all as native ComfyUI components under normal model management. This is what makes this pack different from the renderer-only path, and it's the reason the pack is worth 45 GiB of downloads.

What it actually loads

The planner isn't one model; it's three parts that stay independently offloadable so ComfyUI can shuffle them in and out of VRAM:

• the Qwen2.5-VL language + vision stack,
• the Bernini connector that bridges language output to the planning heads,
• the VIT flow decoder that produces the plan's visual conditioning.

Because they're ordinary ComfyUI patched modules rather than a hidden Diffusers pipeline, offload, low-VRAM mode, and the standard memory manager all work on them. That's the whole design pitch of this pack, and the loader is where it starts.

The two inputs, that's it

This is a loader, so it's the simplest node in the pack:

• repack_manifest - a dropdown over the repack-manifest.json files found under ComfyUI/models/bernini_v2/. Pick the one inside the Bernini-v2-balanced-int8/ (or Bernini-v2-bf16-native/) folder. It drives where the loader looks for the planner's shards.
• dtype - bfloat16 or float16. Leave it on bfloat16; the Balanced INT8 package stores weights as INT8 but computes in BF16, so this setting is about compute precision, and bfloat16 is where the tests were run.

Output is a planner object that feeds straight into Bernini v2 Plan. There's no other consumer for it, and nothing else in the workflow touches it directly.

Practical notes

• Give it a sec. This is a big load. The first time the plan runs, you'll see Qwen and the VIT decoder stream into memory. That's normal; the whole point is that they come back out afterward.
• The manifest matters. If the dropdown is empty, the models aren't where the pack expects them - the bernini_v2 folder under ComfyUI's model root is the home, not models/checkpoints and not your desktop. Re-download with the README's hf download commands and the folder structure is right.
• This node is why the pack exists. The renderer side (Wan loader, guider, UniPC sampler) you can get from other Bernini setups. This loader - native Qwen + VIT planning under ComfyUI memory management - you can't. If you only need quick edits, the renderer-only path is still a lighter download; if you want the full model, this is the door.

Installing it

One pack, one install. ComfyUI Manager: search "Bernini v2 (Native)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-bernini-v2-T8.git

Then restart ComfyUI (needs 0.33+). Models from t8star/Bernini-V2-Comfy:

hf download t8star/Bernini-V2-Comfy --include "Bernini-v2-balanced-int8/*" --local-dir C:/path/to/ComfyUI/models/bernini_v2

The Balanced INT8 package is ~45.6 GiB (the BF16 reference is 83 GiB - get INT8). Also grab wan_2.1_vae.safetensors into ComfyUI/models/vae/. Then load the example t2v.json workflow and hit run.