moritalous/Ministral-3-3B-Base-ERNIE-Image-TE-sdcpp-GGUF

Ministral-3-3B-Base text encoder for ERNIE-Image — stable-diffusion.cpp GGUF

Unofficial GGUF conversion of the text encoder used by ERNIE-Image and ERNIE-Image-Turbo, prepared for stable-diffusion.cpp (--llm). Not affiliated with Baidu or Mistral AI.

Files

What was changed

The ERNIE-Image text_encoder/ is a Mistral3Model whose language model is byte-identical to mistralai/Ministral-3-3B-Base-2512 (spot-checked tensors match). ERNIE-Image uses the second-to-last hidden state of that language model. For stable-diffusion.cpp:

1. 1.only the language-model tensors (language_model.model.*) were kept; the vision tower and projector were dropped,
2. 2.tensors were renamed to the text-only model.* layout that stable-diffusion.cpp expects,
3. 3.weights were quantized with sd-cli -M convert --type {q4_K,q6_K,q8_0} (stable-diffusion.cpp master-899-28b454b).

No weights were fine-tuned or otherwise modified.

Usage (stable-diffusion.cpp)

# diffusion model: unsloth/ERNIE-Image-Turbo-GGUF, VAE: baidu/ERNIE-Image (vae/)
sd-cli \
  --diffusion-model ernie-image-turbo-UD-Q4_K_M.gguf \
  --vae diffusion_pytorch_model.safetensors \
  --llm ministral-3-3b-base-text-Q6_K.gguf \
  --diffusion-fa --steps 8 --cfg-scale 1.0 -W 512 -H 512 \
  -p "a friendly robot mascot" -o out.png

These files use stable-diffusion.cpp tensor naming and may not load in other runtimes (ComfyUI, llama.cpp).

License

Apache License 2.0, same as the original weights (baidu/ERNIE-Image, mistralai/Ministral-3-3B-Base-2512). See LICENSE. Modifications: tensor subset/renaming and quantization as described above.