CIVITAI / Workflows

Ideogram 4 Official Image Generation Workflow

Watch the full video first if you want to understand how this Ideogram 4 official image generation workflow works in practice. The video shows how Ideogram 4 can be used inside ComfyUI for structured poster design, typography-heavy visuals, brand-style images, collage compositions, and layout-controlled image generation. This ComfyUI workflow is designed as a clean Ideogram 4 image generation route. Its main purpose is to provide a stable official-style output pipeline for creators who want to test Ideogram 4 without building a complicated multi-branch graph. The workflow focuses on model loading, prompt encoding, dual-model guidance, quality preset control, resolution handling, sampling, decoding, and final image export. The workflow is built around ideogram4_fp8_scaled.safetensors as the main generation model. It also uses ideogram4_unconditional_fp8_scaled.safetensors as the unconditional branch, qwen3vl_8b_fp8_scaled.safetensors as the Ideogram 4 text encoder, and flux2-vae.safetensors as the VAE. These are the core files required for local deployment. The graph also includes CLIPTextEncode, ConditioningZeroOut, CFGOverride, DualModelGuider, Ideogram4Scheduler, RandomNoise, KSamplerSelect, EmptyFlux2LatentImage, SamplerCustomAdvanced, VAEDecode, and SaveImage. The most important feature of this workflow is its structured prompt design. Ideogram 4 is especially strong for images that need readable text, graphic layout, poster composition, logo-like visuals, magazine-style design, and clear object placement. Instead of relying only on a loose natural-language prompt, this workflow encourages JSON-style captions. The JSON prompt can define the high-level description, background, visual style, lighting, color palette, and composition elements. This gives the model clearer design intent and makes it more useful for cover images, YouTube thumbnails, Chinese posters, advertising visuals, and social media key art. The dual-model guidance structure is another key part. The main Ideogram 4 model reads the prompt and tries to follow the target image direction. The unconditional model provides a baseline. DualModelGuider compares the two branches and pushes the output toward the prompt target. CFGOverride then controls how guidance is applied during sampling, helping the workflow stay aligned without making the image overly rigid. The workflow also includes a q