Make and edit images locally with LLaDA-Image: choosing Base or Turbo

For a new illustration, check whether the subject reads clearly and leaves room for a title. For product-photo editing, focus on details that must not change, such as the cup logo and drink color. Do not judge both outputs with one generic ‘looks good’ score; define what each task must preserve.

For an illustration, check that there is room for a title rather than asking the image to render exact lettering. For editing, specify what to change and preserve in one instruction, then compare against the original for changes to logo text, cup rim, or fingers.

The official examples use 50 steps for Base and 4 for distilled Turbo. Use Turbo drafts to explore color and layout, then submit a preferred prompt to Base as well to compare composition and detail.

Total time also includes model loading, text processing, resolution-dependent computation, and file saving. Record both sampling time and end-to-end time. Compare whether composition, color, and shadows meet your needs, changing one condition at a time when adjusting the result.

For reference-photo editing, state both what should change and what should remain. For example, request a bright indoor background while keeping the cup logo and drink color. Change only the background first, then add another element such as a shadow if needed; this makes changes easier to trace.

The official pipeline²'s editing mode takes an input image and output size; each dimension must be a multiple of 32. For a size error, check that requirement. If the size is accepted but the logo changes, inspect the instruction and content preservation separately. Passing the size check and preserving source details are different tests.

The official repository provides a Diffusers pipeline and Base and Turbo checkpoints³. Its example environment uses PyTorch⁴ with Python 3.11 and CUDA⁵. Select text mode for new-image generation and editing mode for reference edits; record checkpoint, precision, size, and step count.

A ComfyUI workflow may differ from simply translating official commands into a graph. Check required nodes, filenames, and application versions, and save the first working graph as a baseline. Weights fitting on disk does not establish enough memory for the text encoder, image generator, intermediate data, and resolution. Test repeated generation and editing with your usual apps open.

Run the same request three times, separating initial loading from subsequent generation and editing. Compare the middle result to reduce the influence of a single temporary delay. Judge illustrations by the Turbo draft's composition and mood, and compare product photos with the original to check the logo and drink color.

Custom nodes execute code, so verify their source and version. The site's wait-time experience simulates results by device; it is not a speed or quality measurement of LLaDA-Image. If the required generation and editing run reliably alongside your usual apps, an immediate upgrade may not be necessary. Compare candidate hardware under the saved graph's conditions when repeated waiting or memory limits actually block work.