Fal Reference-to-Image

Generating the same subject across images is the problem this ecosystem has never fully solved locally - the IP-Adapter stack that used to be the answer is now SDXL-era legacy, and the newer bases want the reference image fed into the model's context instead. Fal Reference-to-Image is the cloud shortcut: one node, up to four reference images, and a dropdown of the fal endpoints built for subject-reference and multi-reference generation - the IP-Adapter-style jobs, hosted.

It sits in the ComfyUI-Fal-Gateway pack as the reference-version of the image-to-image node, sharing its image_count counter UX with the video nodes. You're not running an adapter on your card; the cloud model looks at your references while it generates, which is a fundamentally different mechanism from injecting adapter embeddings.

Inputs and outputs

• image_count - INT counter, 1 to 4, +/- arrows. Dial it up and more image sockets appear.
• image_1 - required reference. image_2 .. image_4 - optional, revealed by image_count. Wire LoadImage into each.
• model_id - dropdown of the reference-capable image endpoints.
• prompt - what you want generated given the references. Subject-persistent phrasing ("same character, new pose") works better than attribute spam.

Outputs: image (IMAGE tensor → SaveImage), image_url (STRING), info (STRING).

How it works

Same machinery as the rest of the pack: your reference tensors are uploaded to fal's CDN, collected into the multi-image parameter the endpoint's schema expects, and the result is downloaded and decoded back into a tensor. The image_count frontend shows or hides sockets so "how many references" is a dial, not a rewiring job. The honest ceiling, same as everywhere: reference consistency is good but not perfect, and it's the multi-subject case where every approach - local or cloud - still stumbles.

Install

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/modbender/ComfyUI-Fal-Gateway
pip install -r requirements.txt
export FAL_KEY=your_fal_key

Restart ComfyUI; the nodes are under Fal-Gateway. No adapters, no model downloads - the reference handling happens on fal's side.

Gotchas

• Per-call pricing, and reference models bill on the pricier side. The title-bar cost badge gives you the estimate; check it before queueing a sheet of variations.
• Wire enough references for the model. Multi-reference endpoints that require N inputs will error with a "required image input not connected" style message if you under-wire - add the reference rather than fighting the count.
• Fresh nodes can show <no models available> briefly while the catalog fetch warms; right-click → "Fal-Gateway: refresh catalog cache" if it lingers.
• No key → "FAL_KEY not set" hard error.

Honest take: if you're on SDXL or Illustrious with a working IP-Adapter setup, this node is a paid alternative to a free local stack - and probably not worth it. Where it earns its keep is when you're on a 2026 base where the reference-image workflow means a big edit model and you want the closed, hosted version of that instead of training a LoRA or fighting local VRAM.