This guide turns Aiconomist’s 9:59 tutorial into a practical build sheet. The core idea is to stop hand-writing every Ideogram prompt field: use a local Ollama/Gemma-style LLM to inspect a reference image, generate the structured Ideogram JSON prompt, then use bounding boxes and optional LoRAs to control composition, product details, realism, and character consistency.
The video’s upgraded workflow combines four pieces:
An Ollama/Gemma-style local LLM reads a reference image and user override prompt, then fills the Ideogram 4 prompt-builder fields automatically.
You draw boxes over the target character, background, phone, jersey, logo, or other details so Ideogram knows where each element belongs.
The presenter uses FP8 models, but notes that a GGUF quantized version exists for lower-VRAM setups.
Use aesthetic LoRAs such as the “Lenovo/Lora Nova” analog-realism look, or a custom character LoRA trained on authorized images, to improve consistency.
In the transcript, the automated prompt fields are filled by the Gemma/Ollama node, while the presenter leaves the photo style field manual so it can be quickly changed between candid smartphone photo, cinematic shot, DSLR photo, and other looks.
If you have enough VRAM, use the FP8 workflow/model path shown by the presenter. If not, use the GGUF quantized version of the workflow, then confirm the nodes point to the GGUF loader/model files instead of FP8 files.
Open ComfyUI Manager and update the Ideogram/KJ prompt-builder nodes. The video specifically notes that the prompt builder gained official support for a background image input, which appears at the bottom of the bounding-box draw interface.
For Instagram-style images and carousels, the presenter uses 3:4 aspect ratio at 2 megapixels. For faster tests or lower VRAM, use 1 megapixel until the prompt/boxes are dialed in.
The text fields are generated automatically, but keep the high-level photo style field manual if you want fast creative control. Examples from the video: candid smartphone photo, cinematic shot, DSLR-style photo.
The user prompt should tell the LLM what to change from the reference, not merely repeat what is already visible. Example from the video: change the subject into a 40-year-old Italian woman with dark brown hair.
Boxes work best as brief, localized instructions: “black iPhone 16 in hand,” “Spanish national football jersey, one star above crest,” “red nail polish,” or “backseat car interior.” Avoid long paragraphs inside a box; reserve broader style and identity instructions for the main user prompt.
The presenter enables an analog-realism LoRA after a baseline image. The result is softer skin texture, more natural lighting, and less plastic-looking AI skin. Use this when the base output is too synthetic.
A custom Ideogram 4 character LoRA can preserve identity across poses and outfits. Put the LoRA trigger word directly inside the character box for stronger adherence.
| Problem | Likely cause | Fix |
|---|---|---|
| Prompt fields are not filling automatically | Ollama model not running, model name mismatch, or splitter node not connected | Run a small LLM-only test, confirm the selected Ollama model is installed, then check the connection from LLM output to field splitter/prompt builder. |
| Bounding-box node does not show the reference/background image | Outdated KJ/Ideogram prompt-builder nodes | Update custom nodes in ComfyUI Manager and restart ComfyUI. |
| Out of VRAM | Image model, LLM, LoRAs, and high resolution are competing for memory | Use GGUF/quantized model path, lower to 1 MP, unload the LLM after prompt generation, disable LoRAs while testing, or generate at lower batch size. |
| Pose or composition is close but not ideal | Seed variation needed | Change the seed and regenerate before rewriting all prompts. |
| Logo/team detail is wrong | Small details are underspecified or not boxed | Add a dedicated box over the detail, describe star count/crest/brand precisely, and inspect before posting. |
| Face looks right but expression is wrong | Character LoRA is active but expression cue is weak | Add “smiling,” “neutral expression,” or the intended expression directly in the character box. |
| Skin looks plastic | Base model/style is too synthetic | Enable the analog-realism LoRA after the composition works; do not use it as the first debugging variable. |
This guide is based on the provided video transcript and description links. Exact node/model filenames may change; verify against the downloaded workflow and current model/custom-node pages before installing.