CIVITAI / Workflows

Ideogram 4 Structured JSON Image Reconstruction Workflow

Watch the full video first if you want to understand how this Ideogram 4 structured JSON image reconstruction workflow works in practice. The video shows how a reference image can be analyzed, converted into an Ideogram 4 JSON prompt, and then regenerated through a structured image generation pipeline. This ComfyUI workflow is designed for Ideogram 4 image reconstruction through structured JSON prompting. Its main purpose is not ordinary text-to-image generation. Instead, the workflow starts from a reference image, analyzes its visible composition, and rebuilds the image as an Ideogram 4-compatible structured JSON prompt. This makes it especially useful for image style reconstruction, layout recovery, poster remaking, visual reference rebuilding, typography layout testing, and design-oriented “image washing” workflows. The key idea is simple: the user does not need to manually write a long prompt. The workflow uses a reference image as the main input. The visual analysis chain observes the image, extracts the subject, background, composition hierarchy, visible text, color system, lighting, medium, and layout relationship, then generates a structured JSON prompt. That JSON prompt is sent into the Ideogram 4 generation route, allowing the model to recreate the design logic instead of only guessing from a loose natural-language description. The workflow is built around Ideogram 4 image generation. It uses ideogram4_fp8_scaled.safetensors as the main model, Flux2 VAE for decoding, Ideogram4Scheduler for sampling control, DualModelGuider for guided generation, CFGOverride for guidance behavior, EmptyFlux2LatentImage for canvas creation, SamplerCustomAdvanced for final denoising, VAEDecode for image decoding, and SaveImage for final export. The workflow also includes a structured prompt encoding section where the generated JSON is passed into CLIPTextEncode. The most important part is the image-to-JSON reconstruction section. The workflow note describes the main route as LoadImage → image_scale_pixel_v2 → RHLLMChatNode image1, combined with a system prompt and target width / height. RH visual completion observes the reference image and outputs an Ideogram 4 JSON prompt. The width and height are used to help plan bounding boxes, layout proportions, and composition scale. This makes the prompt more useful for structured regeneration, especially when the original