---
license: apache-2.0
base_model:
  - vllm-sr/d3-flash
  - vllm-sr/d3-mini
  - vllm-sr/d3-nano
  - vllm-sr/d3-lite
library_name: diffusion-single-file
tags:
  - comfyui
---

# d3 (ComfyUI)

The backbones of [d3-flash](https://huggingface.co/vllm-sr/d3-flash), [d3-mini](https://huggingface.co/vllm-sr/d3-mini), [d3-nano](https://huggingface.co/vllm-sr/d3-nano) and [d3-lite](https://huggingface.co/vllm-sr/d3-lite) by vllm-sr as single files for ComfyUI. The weights are unchanged except for the format.

| File | Format |
| --- | --- |
| `d3_{flash,mini,nano,lite}_bf16` | bf16, as released |
| `d3_{flash,mini,nano,lite}_int8_convrot` | int8 ConvRot |

Each file also carries its size's decision head: the readout (`d3_readout.weight`, float32) and `decision_config.json` (metadata `d3_decision_config`), unchanged from the original repository. The bf16 files otherwise only rename the keys to ComfyUI's Qwen3.5 layout (`model.` prefix). In the int8 files, only the language-model linear layers are quantized, in ComfyUI's `comfy_quant` format (ConvRot group 256 with per-row scales). Embeddings, norms and the vision encoder stay bf16.

Put one file in `ComfyUI/models/text_encoders` and use it with [ComfyUI-TypedDecision](https://github.com/nomadoor/ComfyUI-TypedDecision): pick `d3` as the adapter and the file as the backbone. Nothing else is needed.