Ideogram V3 Remix

Remix takes an image and a strength value and produces a variation. If that sounds like img2img with extra steps, the difference is the DESIGN style option and the model behind it: Ideogram's hosted v3 is the generation that made it the benchmark for legible text in generated images - logos, packaging, posters, anything where lettering has to actually read.

That's the case for paying per call here rather than running something local. A local img2img pass will happily give you variation; it will not reliably give you a legible wordmark. The default prompt for this node - "Change the cube into a sphere" - tells you how small a change the API is happy with.

Inputs

prompt (required, up to 5000 characters) is the remix instruction. Keep it tight when strength is high and specific when it's low.

image_url (required) is the source image as an IMAGE socket. That's the only required image; everything else is optional.

strength runs from 0.01 to 1, defaulting to 0.8 - the same concept as denoise. Low values hold the original composition and just restyle; high values let the model rebuild more of the frame from your prompt.

style is AUTO (default), GENERAL, REALISTIC, or DESIGN. DESIGN is the interesting one for graphic work; REALISTIC for photographic remixes. rendering_speed is BALANCED (default), TURBO, or QUALITY - the classic draft/final split, and worth using as such, because Turbo iterations are what make this affordable.

expand_prompt (default true) is MagicPrompt, Ideogram's rewriter. For a design task where you're pinning down exact wording, turn it off; for a vague idea, let it work.

image_size is square_hd (default), square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9 - six buckets, not free dimensions. num_images is 1–4; only the first lands on the image output, the rest are in all_urls_json. seed for reproducibility attempts.

negative_prompt defaults to blurry, low quality, distorted, watermark and is capped at 5000 characters. It only helps where your positive prompt doesn't already contradict it - Ideogram's positive prompt wins conflicts.

timeout_seconds (1200 default) and callback_url finish the list.

Outputs

image is the native result. Then url, all_urls_json, task_id, raw_json, credits_consumed and credits_left. Images are archived to ComfyUI/output/KIE-Results/images/ under content-hashed filenames, so consecutive runs don't overwrite each other.

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. Dependencies: requests, Pillow, numpy, PyYAML. No model files - this is a hosted call, which also means your image leaves your machine. Windows one-click installer in dist/ if you'd rather not use a shell. Manager may or may not list the pack yet; the README still describes the registry route as forthcoming.

Set the key once in Settings → KIE.ai Nodes Next → Connection. It's validated against KIE before the save, returns only a masked fingerprint to the UI, and is absent from workflow JSON. Headless installs: KIE_API_KEY, which takes priority over the saved key.

Notes from using this class of node

Ideogram 4 changed the local picture. The 4.0 weights are downloadable now - 9.3B, non-commercial licence, a safety filter trained in, and a prompting style built on JSON captions with bounding boxes. Local, free per image, and fussy about prompts. This v3 API node is the opposite trade: prose prompts, no setup, per-call billing. Know which one you want.

Strength above ~0.9 stops being a remix. You're regenerating with the source as a hint, and small details - including any lettering - will move.

Four images is four charges. num_images is a multiplier and the node's credits_consumed output will say so out loud.

Timeout means the task continues and the credits are gone. Grab task_id from the error and check with the Task Status node.

Rerunning identical inputs returns the cached image without submitting. Change strength or prompt to spend.