README.md

---
license: apache-2.0
base_model: ByteDance/Bernini-R
pipeline_tag: text-to-video
library_name: gguf
tags:
  - gguf
  - wan2.2
  - comfyui
  - bernini-r
  - text-to-video
  - image-editing
  - video-editing
---

# Bernini-R — GGUF (high & low noise experts)

**[💻 Code / ComfyUI nodes](https://github.com/neuregex/ComfyUI-BerniniR)**
 > **[💬 Discord — updates, roadmaps, projects, or just to chat](https://discord.gg/HxfP9TnctJ)**
 > **[🧬 Bernini-R](https://huggingface.co/ByteDance/Bernini-R)**
 > **[Wan2.2-T2V-A14B](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B)**

GGUF quantizations of **[ByteDance/Bernini-R](https://huggingface.co/ByteDance/Bernini-R)**
(= Wan2.2-T2V-A14B + source-id RoPE + multi-condition APG guidance), for use in ComfyUI with
**[ComfyUI-BerniniR](https://github.com/neuregex/ComfyUI-BerniniR)** + [`city96/ComfyUI-GGUF`](https://github.com/city96/ComfyUI-GGUF).

Bernini-R is a **dual-expert** model (Wan 2.2): a high-noise expert sets the composition and a
low-noise expert refines detail. Both are quantized here. GGUF carries no fp8 tensors, so the two
experts coexist in **24 GB VRAM** without the offload crash the fp8 build hits.

## Files

| File | Expert | Quant | Size |
|------|--------|-------|------|
| `bernini_r_high_noise_14B-Q4_K_M.gguf` | high-noise | Q4_K_M | 9.66 GB |
| `bernini_r_high_noise_14B-Q5_K_M.gguf` | high-noise | Q5_K_M | 10.8 GB |
| `bernini_r_high_noise_14B-Q8_0.gguf`   | high-noise | Q8_0   | 15.4 GB |
| `bernini_r_low_noise_14B-Q4_K_M.gguf`  | low-noise  | Q4_K_M | 9.66 GB |
| `bernini_r_low_noise_14B-Q5_K_M.gguf`  | low-noise  | Q5_K_M | 10.8 GB |
| `bernini_r_low_noise_14B-Q8_0.gguf`    | low-noise  | Q8_0   | 15.4 GB |

`Q5_K_M` is the recommended balance; `Q8_0` for best quality, `Q4_K_M` for the lowest VRAM.

## Usage in ComfyUI

1. Install **[ComfyUI-BerniniR](https://github.com/neuregex/ComfyUI-BerniniR)** and **[ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF)**.
2. Put the `.gguf` files in `ComfyUI/models/unet/`. You also need the Wan VAE (`wan_2.1_vae.safetensors`) and the UMT5 text encoder (`umt5_xxl_fp8_e4m3fn_scaled.safetensors`).
3. **t2v / t2i** (`source_id=0` is identical to stock Wan 2.2): one `UnetLoaderGGUF` → your sampler.
4. **Editing (i2i / v2v), both experts:** load each GGUF with `UnetLoaderGGUF`, send each through
   **BerniniR · Apply Patches**, then into **BerniniR · Guider** (`model` = high, `model_low` = low).
   The guider switches expert by timestep (t=875) and runs the APG guidance.
   Ready-made graph: `workflows/ui/bernini_i2i_gguf_dual.json` in the node repo.

License: Apache-2.0 (same as the base model).