Universal Conditioner adapters
Adapters for ComfyUI-UniversalConditioner: one language model on your own server works as the text and vision encoder for several image and video models in ComfyUI, and a small adapter per model maps its hidden states into that model's conditioning space. ComfyUI loads no text encoder.
Adapters
| Target model | File | License | Card |
|---|---|---|---|
| Qwen-Image 2.1 | qwen-image-2.1/unicond-qi21-v3.2-seq.safetensors (269 MB) |
Qwen Research License, non-commercial use only | qwen-image-2.1/README.md |
Each adapter carries the terms of its target model, in its own folder. Read the LICENSE and NOTICE beside the file before you use it. The Qwen-Image 2.1 adapter is built with Qwen.
Use
- Install the ComfyUI-UniversalConditioner node pack.
- Put the adapter file in
ComfyUI/models/universal_conditioner/. - Keep the target model's text encoder file in
ComfyUI/models/text_encoders/. It is read from disk for its token table and is not loaded onto the GPU. For Qwen-Image 2.1 that isqwen3vl_8b_int8_convrot.safetensors. - Run a server that answers
POST /v1/hidden_stateswith the source model loaded, and set its address on the Universal Conditioner node. The pack's README describes the request and response.
Source model
The adapters read hidden states from Qwen3.8-27B.
- Recommended: neroued/Qwen3.8-27B-nvfp4-NInfer. The Qwen-Image 2.1 adapter is fitted on this model.
- Also supported: igorls/Qwen3.8-27B-Uncensored-NVFP4-NInfer. Scores on it are slightly lower; the adapter card lists both.
Status
Experimental. The adapter card gives the held-out scores, the side-by-side checks against the native encoder, and the known limits.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support