Ideogram - Character Remix

Remix is the "same picture, different picture" node. You give it an image and a prompt, and it generates a variation - how far the variation drifts is set by strength, which is the same idea as denoise in a local img2img graph. Low strength nudges, high strength reinvents. Ideogram's hosted v3 remix also takes character references, which is what separates it from a plain image-to-image.

Reach for it when you have a shot you like and want the same subject in a new pose, framing or light without re-rolling the face. Given how much of 2026's character-consistency advice boils down to "generate once, then edit the image," this is that pattern with a slider on it.

Inputs

prompt (required, up to 5000 characters) describes what you want to change or become. The node's default is a long, dense paragraph - that density is the house style here; Ideogram responds well to a full scene description rather than a keyword list.

image_url (required) is the base image as an IMAGE socket. There's also an optional image_urls IMAGE input for additional images - two similarly-named fields that do different jobs, so check which socket you're plugging into.

reference_image_urls is the character-reference IMAGE socket: the images that define who needs to stay the same. reference_mask_urls is a STRING for masks applied to those character references - and per its tooltip only the first mask is used, the rest are ignored, with a 10MB total budget across references. It takes a URL, not a MASK socket.

strength defaults to 0.8 and runs from 0.1 to 1. This is the dial that matters most. Start around 0.5 if you're protecting the composition, go to 0.9+ if you're essentially re-rendering and only keeping the subject.

rendering_speed: BALANCED (default), TURBO, QUALITY. style: AUTO, REALISTIC, FICTION. expand_prompt (default true) is MagicPrompt doing its rewriting thing - off when your prompt is deliberate. image_size is square_hd (default), square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9. num_images is 1–4, and remember only the first comes back on the image output; the rest are in all_urls_json.

negative_prompt here is capped at 500 characters per its tooltip - much tighter than the 5000 on the text-to-image node, so no pasting a novel into it. seed for reproducibility attempts, timeout_seconds (1200 default) and callback_url for plumbing.

Outputs

image (native IMAGE), plus url, all_urls_json, task_id, raw_json, credits_consumed and credits_left. Results are archived under ComfyUI/output/KIE-Results/images/.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt

Restart ComfyUI. requests, Pillow, numpy, PyYAML; no weights to download, because these nodes call the hosted API. (Ideogram's 4.0 weights are downloadable now under a non-commercial licence - that's a separate local setup, not this node, and its prompting style is JSON, not prose.) Windows one-click installer in dist/.

Key once: Settings → KIE.ai Nodes Next → Connection. It's validated before replacing a working key and only a masked fingerprint is shown afterwards. KIE_API_KEY for headless installs, and it wins over the saved key.

Practical notes

Strength is not a quality slider. 0.8 is not "better" than 0.4; it's "further from your input." The useful move is to run the same prompt at two or three strengths and pick.

num_images bills per image. The credits_consumed float on the node will tell you exactly what a 4-image run cost, which is the honest way to discover this.

Character drifts anyway. A reference is a strong hint, not a lock - the pack author says so, and the community's reports on reference-based consistency agree. If you need the same face across dozens of images, a trained LoRA remains more reliable, and you can't train one on a hosted-only model.

Timeout errors are already paid for. If the node gives up waiting, the job is still running on KIE's side. Use the task_id from the error with the Task Status node.

Identical rerun = cached output, no charge, same image. Change strength or the prompt to actually submit.