arkennemasis fal · MiniMax Image 01 Subject Reference · $0.01/image

Character consistency is the problem that never got solved, only rerouted - and one of the real routes is "hand the model a reference image and let it look at the thing while it draws." That's this node. Feed it a portrait and a prompt, get a new scene with the same subject, at a penny an image.

Where this fits in 2026's consistency stack

The KB's honest summary: the community's answer moved from "encode the character and inject it at generation time" (IP-Adapter, FaceID, PuLID, all still functional) to "generate once, then edit that image" with an instruction-edit model. LoRAs still work; people just stopped recommending them for a character you'll draw ten times.

A reference-image generator sits between those poles. It's cheaper and lazier than an edit chain - no mask, no crop-and-stitch, no drift compounding across a chain of edits - but it's also the model's interpretation of similarity rather than a guaranteed pixel-level identity. For a storyteller who needs a face to read as the same person across slides, storyboards and product shots, one cent per attempt is a very forgiving way to find out how close it gets. For a paid campaign where the face is the product, you'd still want the edit-model route and a human checking every frame.

How it works

Standard pack mechanics over fal's fal-ai/minimax/image-01/subject-reference endpoint: read FAL_KEY from the .env beside the launcher, validate, estimate cost against max_cost_usd, upload your reference image to fal's CDN (free), submit one billed request, poll, download into output/fal/fal-ai_minimax_image-01_subject-reference/. Hailuo/Image-01's owner MiniMax is closed and API-first, so there's no local equivalent to fall back to - this is a hosted capability, priced per image.

Inputs

• image - the subject reference, and the whole reason you're on this node instead of the plain Image 01. Use a clean, well-lit, front-on shot. The tooltip is explicit that this drives "consistent character appearance"; a busy or stylised reference gives the model more to be confused by.
• prompt - up to 1500 characters, describing the new scene and action while not re-describing the subject's face. Let the reference do that job.
• aspect_ratio - 1:1, 16:9, 4:3, 3:2, 2:3, 3:4, 9:16, 21:9.
• num_images - 1 to 9, and the cost estimate scales with it.
• prompt_optimizer - default off here.

Optional: max_cost_usd ($20 default; the run halts before uploading if the estimate exceeds it, 0 = uncapped), max_concurrent (default 1 - one paid fal call at a time across the run), reuse_identical_run (default on; identical inputs return the file you already paid for), extra_json (raw JSON merged into the request for unexposed fal fields).

Outputs: images - the returned pictures batched onto one IMAGE output - plus image_paths and info (request id, estimate, saved files, fal's response). Wire it into a Save node, or into the pack's Option Board if you want a whole variation run laid out on one sheet to judge.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install httpx

Manager route: Install via Git URL with the repo. Then the key:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

in the .env beside run_nvidia_gpu.bat. Restart and look under arkennemasis/fal/Image/MiniMax.

Practical notes

Two things make this node work better than it has any right to at one cent: batch width and a fixed reference. Set num_images to 9, keep the same reference image, and vary only the scene description - you're buying nine interpretations of the same person for nine cents, which is a reasonable way to find the one that holds up. Changing the reference between attempts is what makes consistency look broken, not the model.

Also worth knowing: penny images are a draft tier, not a final tier. Generate nine, pick one, then upscale locally rather than paying for more attempts. And if the same run keeps returning the exact same file, that's reuse_identical_run being thrifty - turn it off when you want a fresh sample from an unchanged prompt.