fszontagh/LLaDA-Image-GGUF

LLaDA-Image GGUF quantized files

Quantized transformers, text encoders and pre-merged connector files for running the 50-step LLaDA-Image base model with stable-diffusion.cpp.

The license of the quantized files follows the license of the original model:

• LLaDA-Image: apache-2.0

Everything here belongs to the 50-step base model. The 4-step distilled model, LLaDA-Image-Turbo, has a different transformer, text encoder, QueryFormer and text projection; only the VAE, the SigVQ encoder and the tokenizer are shared, so the two sets cannot be mixed. Turbo files are at fszontagh/LLaDA-Image-Turbo-GGUF.

These files are converted using https://github.com/leejet/stable-diffusion.cpp

Which files do I need?

Every run needs five things: a transformer, a text encoder, a VAE, a connectors file, and an external tokenizer.json. This repo provides the first three; only the VAE comes from upstream.

Transformer

file	notes
llada-image-f16.gguf	the original weights in GGUF form
llada-image-q8_0.gguf	smaller, for tighter setups

Text encoder

Quantized from the upstream 33 GB bf16 shards. The Files tab is the authoritative list.

file	notes
llada-image-text_encoder-q4_K.gguf	smallest
llada-image-text_encoder-q5_K.gguf
llada-image-text_encoder-q6_K.gguf
llada-image-text_encoder-q8_0.gguf	largest

Connectors

LLaDA-Image needs a QueryFormer and a text projection alongside its text encoder, and a SigVQ image encoder for editing. Upstream ships these as separate directories, while --embeddings-connectors takes a single file, so they are merged here with their component name as the tensor-name prefix.

file	use
llada-image-connectors.safetensors	text to image only
llada-image-connectors-edit.safetensors	text to image and image editing (adds SigVQ)

VAE

Not included; it is unchanged from inclusionAI/LLaDA-Image. Use vae/diffusion_pytorch_model.safetensors, referred to below as llada_vae.safetensors.

Tokenizer

Not included. stable-diffusion.cpp does not embed the LLaDA2 vocabulary, so pass tokenizer/tokenizer.json from inclusionAI/LLaDA-Image with --tokenizer. The file is the same for both checkpoints.

Example commands

Text to image:

./bin/sd-cli \
  --diffusion-model llada-image-f16.gguf \
  --llm llada-image-text_encoder-q8_0.gguf \
  --tokenizer tokenizer.json \
  --vae llada_vae.safetensors \
  --embeddings-connectors llada-image-connectors.safetensors \
  -p "a lovely cat holding a sign says 'llada.cpp'" \
  --steps 50 --cfg-scale 5 -W 1024 -H 1024 --diffusion-fa -o output.png

Image editing:

./bin/sd-cli \
  --diffusion-model llada-image-f16.gguf \
  --llm llada-image-text_encoder-q8_0.gguf \
  --tokenizer tokenizer.json \
  --vae llada_vae.safetensors \
  --embeddings-connectors llada-image-connectors-edit.safetensors \
  --ref-image input.png \
  -p "change the sign text to 'sd.cpp'" \
  --steps 50 --cfg-scale 5 -W 1024 -H 1024 --diffusion-fa -o output.png

Notes

• Use 50 steps and --cfg-scale 5 for the base model.
• Edit at 1024x1024. At 512x512 this model returns the reference image almost unchanged instead of applying the instruction.
• Width and height are rounded up to a multiple of 16; editing expects them divisible by 32.
• The weights are large, but stable-diffusion.cpp streams them: --max-vram 3 still works.