fszontagh/LLaDA-Image-Turbo-GGUF

LLaDA-Image-Turbo for stable-diffusion.cpp

Quantized transformers and pre-merged conditioning weights for running LLaDA-Image-Turbo with stable-diffusion.cpp.

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

• LLaDA-Image-Turbo: apache-2.0

Everything here belongs to LLaDA-Image-Turbo. The 50-step base model, LLaDA-Image, ships a different transformer, text encoder, QueryFormer and text projection; only the VAE, the SigVQ encoder and the tokenizer are shared between the two, so these files cannot be mixed with it.

No weights were retrained or altered. The GGUF files are quantized from the original bf16 transformer using stable-diffusion.cpp; the connector files repackage the upstream components into a single file, renaming tensors only.

Support for this model was added in leejet/stable-diffusion.cpp#1968. See docs/llada_image.md for the full usage guide.

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 transformer, the text encoder and the connectors; only the VAE comes from upstream.

Transformer

file	size	notes
llada-image-turbo-f16.gguf	13.08 GB	the original weights in GGUF form; what the docs use
llada-image-turbo-q8_0.gguf	6.95 GB	smaller, for tighter setups

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	size	use
llada-image-turbo-connectors.safetensors	0.75 GB	text to image only
llada-image-turbo-connectors-edit.safetensors	3.35 GB	text to image and image editing (adds SigVQ)

Both work for text to image and produce identical output; the edit file simply also contains SigVQ, so one model load serves both modes.

Text encoder

All are quantized from the upstream 33 GB bf16 shards. More levels are being added; the Files tab is the authoritative list.

file	size	notes
llada-image-turbo-text_encoder-q4_K.gguf	9.69 GB	what the docs use
llada-image-turbo-text_encoder-q5_K.gguf	11.69 GB

To build a different quantization yourself, download text_encoder/ from inclusionAI/LLaDA-Image-Turbo and run:

./bin/sd-cli -M convert -m text_encoder/model.safetensors.index.json \
  -o llada-image-turbo-text_encoder-q6_K.gguf --type q6_K

VAE

Not included; it is unchanged from inclusionAI/LLaDA-Image-Turbo. Use vae/diffusion_pytorch_model.safetensors.

Tokenizer

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

Example commands

Text to image:

./bin/sd-cli \
  --diffusion-model llada-image-turbo-f16.gguf \
  --llm llada-image-turbo-text_encoder-q8_0.gguf \
  --tokenizer tokenizer.json \
  --vae llada_vae.safetensors \
  --embeddings-connectors llada-image-turbo-connectors.safetensors \
  -p "a lovely cat holding a sign says 'llada.cpp'" \
  --steps 4 --cfg-scale 1.0 -W 1024 -H 1024 -o output.png

Image editing:

./bin/sd-cli \
  --diffusion-model llada-image-turbo-f16.gguf \
  --llm llada-image-turbo-text_encoder-q8_0.gguf \
  --tokenizer tokenizer.json \
  --vae llada_vae.safetensors \
  --embeddings-connectors llada-image-turbo-connectors-edit.safetensors \
  --ref-image input.png \
  -p "change the sign text to 'sd.cpp'" \
  --steps 4 --cfg-scale 1.0 -W 1024 -H 1024 --diffusion-fa -o output.png

Notes

• Use 4 steps and --cfg-scale 1.0 for the Turbo model.
• Width and height must be multiples of 16 for text to image, and multiples of 32 for editing.
• The weights total about 16 GB, but stable-diffusion.cpp streams them, so a smaller budget works: --max-vram 3 still produces byte-identical output.
• Editing at 1024x1024 needs --diffusion-fa on a 12 GB card.