HiDream-O1 in ComfyUI: setup and model comparisons

This tutorial builds a ComfyUI workflow for the HiDream-O1 base model, which the presenter describes as open source under an MIT license. An updated dev model existed at recording time, but compatible ComfyUI versions were not yet available. The workflow uses a checkpoint loader, patch seam smoothing and a dedicated latent image node for supported sizes.

The presenter demonstrates text-to-image generation with CFG 3 and 28 steps, then tests reference images, image-to-image editing and masked inpainting. Image and latent upscaling examples reach 2560x2560. Size restrictions matter: the closing style-transfer example uses input dimensions divisible by 32, with 30 steps, CFG 2 and denoise 0.3.

Comparisons with ERNIE-Image and Z-Image focus on prompt adherence, rendered text and image detail. In these examples, the presenter favors HiDream-O1 for object placement and lettering, while criticizing its smooth textures and oversaturated finish. Z-Image refinement improves fur detail and reduces saturation in the demonstrated workflow. None of the models fully satisfies the complex cubist prompt.

A Gemma 4 prompt-generation workflow expands prompts using instructions from the project's GitHub page. The presenter reports that escaped quotes and added commas can appear in generated images, so the resulting prompt may need manual editing.