README.md

---
license: apache-2.0
base_model: ByteDance/Bernini-R-1.3B-Diffusers
pipeline_tag: image-text-to-video
library_name: gguf
tags:
  - comfyui
  - bernini-r
  - wan2.1
  - text-to-video
  - image-editing
  - video-editing
  - gguf
---

# Bernini-R 1.3B — ComfyUI bundle (everything you need)
> **[💬 Discord — updates, roadmaps, projects, or just to chat](https://discord.gg/HxfP9TnctJ)**
> 
>  **[💻 Code / ComfyUI nodes](https://github.com/neuregex/ComfyUI-BerniniR)**
> 
> **[🧬 Bernini-R 1.3B](https://huggingface.co/ByteDance/Bernini-R-1.3B-Diffusers)**
> 
> **[14B GGUF (full quality)](https://huggingface.co/neuregex/Bernini-R-GGUF)**

The **small, accessible** Bernini-R for ComfyUI. This repo is **self-contained** — the
1.3B renderer **plus** the VAE and text encoder it needs — so you can edit images/video on
modest GPUs. It's the lightweight sibling of the [14B GGUF repo](https://huggingface.co/neuregex/Bernini-R-GGUF)
(that one is full quality for 24 GB cards).

Bernini-R 1.3B is **single-expert** (fine-tuned from Wan2.1-T2V-1.3B — no high/low MoE), so it's
tiny (~2.6 GB) and runs almost anywhere. It performs close to the 14B on simpler edits
(style transfer, watermark/subtitle removal, local editing) and lags on the hardest tasks.

## Files

| File | What | Size | Folder in ComfyUI |
|------|------|------|-------------------|
| `bernini_r_1.3B-bf16.safetensors` | renderer (native, full precision) | ~2.8 GB | `models/diffusion_models/` |
| `bernini_r_1.3B-Q8_0.gguf` | renderer (GGUF, near-bf16) | ~1.5 GB | `models/unet/` |
| `bernini_r_1.3B-Q6_K.gguf` | renderer (GGUF, high quality) | ~1.1 GB | `models/unet/` |
| `bernini_r_1.3B-Q5_K_M.gguf` | renderer (GGUF, balanced) | ~1.0 GB | `models/unet/` |
| `bernini_r_1.3B-Q4_K_M.gguf` | renderer (GGUF, smallest) | ~0.8 GB | `models/unet/` |
| `vae/wan_2.1_vae.safetensors` | Wan 2.1 VAE | ~0.25 GB | `models/vae/` |
| `text_encoders/umt5-xxl-encoder-Q5_K_M.gguf` | UMT5 text encoder (**GGUF — used by the workflows**) | ~4.1 GB | `models/text_encoders/` |
| `text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors` | UMT5 text encoder (fp8 — optional fallback) | ~6.4 GB | `models/text_encoders/` |

## Usage in ComfyUI

1. Install **[ComfyUI-BerniniR](https://github.com/neuregex/ComfyUI-BerniniR)** and **[ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF)** (required — the graphs load the text encoder with `CLIPLoaderGGUF`).
2. Drop each file in the folder from the table above.
3. Load the workflow `workflows/bernini_i2i_1.3B.json` (or grab it from the node repo). It's a
   **single-expert** graph — no `model_low`. Put your image, run.
   - Renderer — native (`bf16`): **BerniniR · Load Model (native)**. GGUF: `UnetLoaderGGUF` → **BerniniR · Apply Patches** → **Source Stream** → **Guider** (leave `model_low` empty).
   - Text encoder: `CLIPLoaderGGUF` (`type = wan`) → `umt5-xxl-encoder-Q5_K_M.gguf`. The graphs default to the GGUF encoder because the fp8 `.safetensors` one triggers a Windows / torch-2.8 access violation under memory pressure; the fp8 file is kept only as an optional fallback.

> UMT5 GGUF text encoder quantized by **[city96](https://huggingface.co/city96/umt5-xxl-encoder-gguf)**.

## Need full quality on 24 GB?

Use the **14B** dual-expert GGUFs: [neuregex/Bernini-R-GGUF](https://huggingface.co/neuregex/Bernini-R-GGUF).
Same nodes, just wire both experts into the guider (`model` = high, `model_low` = low).

License: Apache-2.0 (same as Bernini-R and the Wan base).