LLaDA-Image Base — ComfyUI INT8 + GGUF
Quantized LLaDA-Image Base weights for the RealRebelAI LLaDA-Image ComfyUI custom nodes.
Based on inclusionAI/LLaDA-Image, this release is intended for lower-memory ComfyUI inference while retaining the Base model workflow.
ComfyUI Custom Nodes
https://github.com/RealRebelAI/LLaDa-Image_ComfyUI
Install into:
ComfyUI/custom_nodes/ComfyUI-LLaDA-Image/
Restart ComfyUI after installation.
Upstream Model
Official Base model:
https://huggingface.co/inclusionAI/LLaDA-Image
Official project:
https://github.com/inclusionAI/LLaDA-Image
This is an unofficial quantized derivative and is not affiliated with inclusionAI.
Included Weights
Base INT8 Transformer
LLaDA-Image-Base-transformer-INT8.safetensors
Place in:
ComfyUI/models/diffusion_models/
This is a native INT8 Safetensors transformer, not a transformer GGUF.
Base Q4_K_M Text Encoder
LLaDA-Image-Base-text_encoder-Q4_K_M.gguf
Place in:
ComfyUI/models/text_encoders/
The custom runtime uses City96 ComfyUI-GGUF support for the quantized LLaDA2 MoE text encoder.
Requirement
Install City96 ComfyUI-GGUF:
https://github.com/city96/ComfyUI-GGUF
VAE
Place the compatible LLaDA VAE in:
ComfyUI/models/vae/
Then select it in LLaDA Image Loader.
The current custom nodes expose native VAE tiled decoding:
vae_tiling:
On
Auto
Off
For low-VRAM GPUs, On is a good starting point.
Base Recommended Settings
Steps: 50
CFG: 5.0
LLaDA-Image Base is the full model rather than the distilled few-step Turbo variant.
Text-to-Image
LLaDA Image Loader
|
v
LLaDA Image Text to Image
|
v
Save Image
Loader example:
diffusion_model: LLaDA-Image-Base-transformer-INT8.safetensors
text_encoder: LLaDA-Image-Base-text_encoder-Q4_K_M.gguf
vae: your LLaDA VAE
dtype: bfloat16
vae_tiling: On
Start with 50 steps / CFG 5.0.
Native Image Editing
The custom nodes support LLaDA-Image's native image-editing mode.
This is not conventional img2img or denoise-strength emulation. The source image is passed through LLaDA-Image's native:
generation_mode="editing"
using its image-conditioning/SigVQ path.
LLaDA Image Loader -----------+
|
Load Image -------------------+--> LLaDA Image Edit --> Save Image
Example:
Turn the fox into a white arctic fox while preserving the forest composition and realistic photography.
For Base editing, start with:
Steps: 50
CFG: 5.0
Editing width and height must be divisible by 32.
The same Base transformer and text encoder are used for generation and editing. No separate editing checkpoint is required.
Base vs Turbo
Base:
Steps: ~50
CFG: ~5.0
Turbo:
Steps: ~4
CFG: ~1.0
Low-VRAM Notes
- Use the INT8 transformer.
- Use the Q4_K_M GGUF text encoder.
- Enable VAE tiling.
- Use CPU offload when necessary.
- Do not load the original full text encoder alongside the quantized encoder.
Model Placement
ComfyUI/
└── models/
├── diffusion_models/
│ └── LLaDA-Image-Base-transformer-INT8.safetensors
├── text_encoders/
│ └── LLaDA-Image-Base-text_encoder-Q4_K_M.gguf
└── vae/
└── <LLaDA VAE>.safetensors
Important
Do not load the INT8 transformer through a GGUF diffusion loader. It is native INT8 Safetensors.
The .gguf file in this release is the text encoder, not the diffusion transformer.
Credits
- inclusionAI — LLaDA-Image architecture and official model
- Hugging Face Diffusers — pipeline/component infrastructure
- City96 — ComfyUI-GGUF / GGML support
- ComfyUI — node and inference ecosystem
- RealRebelAI — custom ComfyUI integration and quantized release
License
The upstream LLaDA-Image model and components remain subject to their original licenses and terms. Review the upstream license before redistribution or commercial use.