CIVITAI / Workflows

Ideogram 4 Reference Latent Reconstruction Image-to-Image Workflow

Watch the full video first if you want to understand how this Ideogram 4 reference latent reconstruction workflow works in practice. The video shows how a reference image can be analyzed, converted into a structured Ideogram 4 JSON prompt, encoded into latent space, and then regenerated through a controlled image-to-image reconstruction pipeline. This ComfyUI workflow is designed for Ideogram 4 image-to-image reference latent reconstruction. Its main purpose is to rebuild an existing image by combining two control routes: a visual-language JSON prompt route and a reference latent remix route. Compared with a pure text-to-image workflow, this graph gives Ideogram 4 both a structured description of the image and a latent-space reference starting point. Compared with ordinary image-to-image workflows, it is more experimental and more design-oriented, because the final stability comes from both the reference latent and the reconstructed JSON prompt. The workflow starts from a reference image. The image is scaled and prepared through the image scaling section, then encoded into latent space through VAEEncode. This encoded latent is sent into the sampler as the starting latent instead of using a blank EmptyFlux2LatentImage canvas. The older text-to-image empty latent path is intentionally disconnected in this version, because the workflow is focused on reference reconstruction rather than pure generation. A key technical point is the SplitSigmasDenoise stage. The scheduler output is split before entering the sampler, allowing the workflow to run a controlled denoise range over the reference latent. This makes the result behave like a forced latent remix route: it may preserve part of the reference image’s composition and structure, while still allowing Ideogram 4 to reinterpret the image according to the prompt and denoise strength. It is not a classic SD 1.5-style img2img pipeline, so the result can drift depending on denoise settings, prompt strength, and reference complexity. The second major control route is the Vision LLM prompt reconstruction chain. The workflow uses a reference image together with target width and height information, then asks the RH visual completion section to generate an Ideogram 4-compatible structured JSON prompt. This JSON can describe the subject, background, layout hierarchy, bounding boxes, readable text, color palette, lightin