MOSSLiteTE / README.md
cicalooo's picture
Upload README.md with huggingface_hub
2bc4c01 verified
|
Raw History Blame Contribute Delete
1.68 kB
---
license: apache-2.0
base_model:
- Qwen/Qwen3.5-0.8B
- recoilme/moss_te_adapter
tags:
- comfyui
- moss-image
- text-encoder
---
# MOSSLiteTE
ComfyUI-native converted weights of [recoilme/moss_te_adapter](https://huggingface.co/recoilme/moss_te_adapter) plus text-only [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B): a lightweight (~1.8 GB) text encoder for MOSS-Image that replaces the native Qwen3.5-9B encoder.
Use with the ComfyUI node pack **[ComfyUI-MOSSLiteTE](https://github.com/cicalooo/ComfyUI-MOSSLiteTE)**.
## Files
| file | place in ComfyUI |
| --- | --- |
| `text_encoders/mosslite_qwen3.5_0.8b_bf16.safetensors` (Qwen3.5-0.8B, text-only: `mtp.*` and `model.visual.*` dropped) | `ComfyUI/models/text_encoders/` |
| `mosslite/mosslite_te_head_bf16.safetensors` (158M regression head, bf16) | `ComfyUI/models/mosslite/` |
```bash
hf download cicalooo/MOSSLiteTE --local-dir /tmp/mosslite
cp /tmp/mosslite/text_encoders/*.safetensors ComfyUI/models/text_encoders/
mkdir -p ComfyUI/models/mosslite && cp /tmp/mosslite/mosslite/*.safetensors ComfyUI/models/mosslite/
```
## Notes
- Approximation: ~0.92 cosine vs the native 9B encoder (per the original adapter card).
- The ComfyUI port matches the reference implementation (parity 1.0).
- Text-only: no image inputs to the text encoder.
- 1024-token prompt limit (longer prompts are truncated).
- Trained on MOSS-Image-Base; untested on SFT.
## Credits
- [recoilme](https://huggingface.co/recoilme) — original moss_te_adapter
- [Qwen](https://huggingface.co/Qwen) — Qwen3.5-0.8B
- [OpenMOSS](https://huggingface.co/OpenMOSS) — MOSS-Image
License: Apache-2.0 (same as the sources).