LLaDA-Image in ComfyUI: Base, Turbo and Model Setup

What is LLaDA-Image?

inclusionAI released LLaDA-Image on September 4, 2026. Its official model cards describe a 6B unified image-generation and editing family. Base recommends 50 sampling steps; the distilled Turbo recommends 4. Both support text-to-image generation, VQ-conditioned generation, reference-image editing and Chinese–English text rendering. Those are upstream capabilities, not evidence that every third-party workflow exposes every mode. PromptToVisual does not host or run LLaDA-Image.

Base versus Turbo: which installation are you reading about?

Base is the official higher-step option; Turbo is the lower-step variant aimed at faster generation and editing. The RealRebelAI integration linked here targets Turbo. Do not assume that swapping in a Base checkpoint will preserve that custom workflow. Fifty versus four steps is a sampling configuration, not a promise that one setup finishes exactly twelve and a half times faster. Hardware, model loading and other processing also affect elapsed time.

1. Identify the official and community projects

The inclusionAI repository is the upstream implementation. RealRebelAI maintains an independent ComfyUI integration and optimized weights. Its adapter code requires City96 ComfyUI-GGUF for the quantized text encoder. Install the custom node repository under your ComfyUI custom_nodes directory using your usual reviewed custom-node installation method, enable ComfyUI-GGUF, and restart ComfyUI. Follow the current repositories for dependency changes rather than running a copied command that assumes a requirements file exists.

2. Put each component in the right folder

The integration documents transformer weights under ComfyUI/models/diffusion_models/, the GGUF text encoder under ComfyUI/models/text_encoders/, and the VAE under ComfyUI/models/vae/. Its INT8 transformer uses Safetensors; the text encoder uses GGUF. They are different components. The loader also needs supporting pipeline components, so downloading only the transformer is not a complete setup. Select the actual downloaded filenames in the loader rather than relying on names cached in someone else’s workflow.

3. Check the current file list before importing a workflow

The linked Hugging Face file tree contains separate generation and editing workflow JSON files. At review time, its text encoder filename omitted the -v3 suffix still shown in the model-card table. The INT8 transformer name also differed between the file list and documentation. This mismatch is a practical reason to check dropdown selections after import. A missing filename does not necessarily mean a model file is corrupt; the workflow may simply reference an older name.

4. Start with the documented Turbo settings

The community README recommends 4 steps and guidance 1.0. For editing, it describes a loader feeding LLaDA Image Edit, a source from Load Image, and a Save Image output. Editing dimensions must be divisible by 32. The official pipeline distinguishes text, VQ and editing modes; VQ mode generates its image tokens from the prompt and does not take a reference image. Native editing should not be described as an ordinary denoise-slider img2img workflow.

5. Verify a small result before expanding the graph

We have not installed or executed this ComfyUI graph on a GPU, so these are source-based setup instructions, not a tested hardware recipe. Begin with one modest task and preserve the logs. Confirm that the loader selects all required components, the workflow produces an image, and Save Image creates a readable file. Inspect dimensions and content after reopening the export. Only then add more nodes or increase output size. An imported graph without missing-node warnings is not yet evidence that inference works.

An original text and editing check

For a self-run test, request a simple café poster with the exact short heading “OPEN TODAY” and a separate Chinese line “今日营业”. Inspect each character rather than judging only the overall composition. Then edit one visible color while asking to retain the wording and layout. Keep both files and compare letter shapes, object count and alignment. This is an illustrative test brief; no LLaDA output is claimed here. A bilingual capability statement does not guarantee that every phrase, font or dense layout will render correctly.

How to interpret quantized results

Treat a reduced-precision setup as its own configuration. Record the transformer and text encoder filenames, workflow revision, steps, guidance, seed and output dimensions. If you compare it with another configuration, change one major variable at a time. A visually similar output on one prompt cannot establish general equivalence. Conversely, a single failed run could be caused by mismatched components or the graph rather than the underlying model. Keep the error log and input instead of immediately replacing several dependencies.

Is this production-ready?

We have not established production reliability, memory requirements or performance on your hardware. The current cards do not contain the earlier quoted “still validating” sentence, so we do not attribute that wording to the author. The files and documentation already show naming differences. Treat this as a community integration requiring your own validation, not a stable service promise. Check upstream and integration licensing separately before deploying it for other people.

Use a hosted workflow instead?

If your immediate goal is a finished image rather than maintaining a local graph, PromptToVisual offers other listed image models. Selecting one does not run LLaDA behind the scenes. Choose based on the task and displayed cost, then inspect the result. Our image comparison tool can help review two exports from any workflow, while the local text tool can place exact final wording on an image without regenerating the scene.

Sources and next steps

Source check: September 9, 2026. These are independent guides. For actual Images 2.5 outputs, see the linked model pages; the LLaDA setup has not been run here.

Official Base model card →

Official Turbo model card →

Official code →

Community ComfyUI nodes →

Community weights and workflows →

GGUF dependency in adapter code →

Flare generator and real samples → · Sunburst generator and real samples → · Images 2.5 and Sketch · LLaDA ComfyUI setup

Continue with the right tool

AI image combiner · Stitch images · Photo collage · Nano Banana 2 guide · Related composition guide