Text encoders and tokenizers¶

Some newer models don't ship with their text encoder inside the main model file. They need one or more auxiliary model files next to the diffusion model:

• a text encoder (e.g. CLIP-L, CLIP-G, T5-XXL, or an LLM such as Qwen), and
• sometimes an external tokenizer (tokenizer.json).

Easy Diffusion exposes these as the Text Encoder and Tokenizer dropdowns in Image Settings. They are usually left empty, and are only needed for the models that require them.

Text encoders¶

Text encoders turn your prompt into the numbers that the diffusion model works with. Models like Flux, SD3, Qwen Image and MOSS-Image can use a separately-downloaded encoder file.

Where to put the files¶

Put the text encoder file(s) in the models/text-encoder folder:

EasyDiffusion/
└── models/
    └── text-encoder/
        ├── t5xxl_fp16.safetensors
        ├── clip_l.safetensors
        └── Qwen3VL-8B-Instruct-Q4_K_M.gguf

Recognized file types are .safetensors, .sft and .gguf. Refresh the model list with the reload icon next to the Model dropdown after adding files.

How to select them¶

Pick the file(s) in the Text Encoder dropdown in Image Settings. You can add more than one text encoder when a model needs several (e.g. CLIP-L and T5-XXL for some Flux models). Models with a built-in text encoder don't need this, so leave the dropdown empty for them.

Vision files (llm_vision)¶

Some image-editing models use a vision encoder so they can look at the reference image you attach, not just your prompt. These are the mmproj files (e.g. mmproj-Qwen3VL-8B-Instruct-F16.gguf) that accompany a multimodal LLM.

Vision files are handled like any other text encoder: put them in models/text-encoder, then select them in the Text Encoder dropdown alongside the main LLM encoder. They are only used by vision-capable models on the v4 engine, together with the Reference Images input. If you load a vision file for a model that doesn't use one, it is simply ignored.

Tokenizers¶

A tokenizer splits your prompt into the tokens that a text encoder reads. Most encoders carry their tokenizer inside the model file, so you rarely need this.

Some models use a text encoder whose vocabulary differs from the one baked into the model. For those, you must download the matching tokenizer.json and point Easy Diffusion at it. A common example is MOSS-Image, which uses Qwen3.5 9B and needs a tokenizer.json that differs from the embedded Qwen tokenizer.

Where to put the files¶

Put the tokenizer file in the models/tokenizer folder:

EasyDiffusion/
└── models/
    └── tokenizer/
        └── tokenizer_qwen3.5.json

Tokenizer files use the .json extension. Refresh the model list with the reload icon next to the Model dropdown after adding files.

How to select them¶

Pick the file in the Tokenizer dropdown in Image Settings. Leave it empty when the model's text encoder already includes a compatible tokenizer.

If the model still produces garbled output, or the logs mention a tokenizer mismatch, double-check that the tokenizer.json matches the text encoder and model you're using. The tokenizer must come from the same release as the text encoder.

Related pages¶

• The v4 engine: reference images and other v4-only features.
• Flux and other new models: which new models need a separate text encoder.
• Custom models: sharing the models/ folder with other UIs.