MiniMax H3 Image LoRA · Edit

Reference-driven editing with your own adapters

This is the heaviest image node in the pack. Up to nine reference images on autogrow sockets, up to three LoRAs stacked on top, and a prompt that addresses the references by name - <Picture 1>, <Picture 2>, and so on. The syntax is worth reading twice because it's the API's convention, not a UI flourish: the tooltip says to refer to references exactly that way, brackets and all.

Why you'd want it: character consistency across a set, product shots with the object from one reference and the styling from another, or a look you trained yourself and now want applied to an edit rather than a fresh generation. Nine references is enough to build a small style bible and hand it to one node.

The practical appeal over the pre-LoRA H3 edit path is control: a LoRA can lock a face or a texture in ways prompt language never quite manages, and it survives across every run without re-describing it.

What happens on queue

Each connected image is encoded and uploaded - a socket carrying a batch becomes several references, in order - and the LoRA nodes upstream resolve their files (uploaded once from ComfyUI/models/loras, or imported by PopNow from a Hugging Face repo or Civitai page, cached for twelve hours). Then the whole thing goes out as one job and comes back as an IMAGE.

Aspect ratio deserves a note. The default is model default, which means "follow the first reference image". The tooltip says as much, and it's the setting to leave alone unless you're deliberately reframing - there's no other node in the graph enforcing your intended crop.

Inputs, ordered by how much they matter

• prompt - what to change, and what to leave. Name references as <Picture 1>, <Picture 2>. "Keep the jacket from <Picture 2>, put the person from <Picture 1> in the room from <Picture 3>" is the shape that works.
• image_urls - the autogrow group, one to nine. Every image after the first raises the quote, which is the honest reason to be selective.
• loras - up to three H3 LoRAs, each with its own scale. Over 2 GiB gets rejected.
• aspect_ratio - model default or an explicit ratio; default follows your first reference.
• resolution - 1k (about a megapixel) or 2k (about four). 2k roughly doubles the price, so it's a finishing decision.
• output_format - jpeg default, png and webp available.
• seed - fixed by default; a repeated queue reuses the result instead of re-billing.

Outputs: images and task id.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Or Manager > search PopNow > Install, or comfy node install popnow-comfyui. The pack declares no dependencies - it rides on what ComfyUI already ships, including the V3 node API - so there's no pip step and nothing to download. Key from popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY before launch, then restart so the catalog loads.

Where it goes wrong

The LoRA is the wrong family. H3 image LoRAs only. Not H3 video LoRAs, not Krea or Flux files, not an SDXL character LoRA. Architecture-bound is the rule everywhere and it bites hardest in a node advertising three slots - it's easy to fill all three with files that do nothing.

Reference numbering. Prompt says <Picture 1>; the sockets are ordered by position on the node. Reorder the wires and the prompt is now describing the wrong image, confidently. Wire, then write.

Nine references as an excuse not to decide. Each extra image raises the quote and dilutes attention. Two or three well-chosen references usually beat nine.

A gated Civitai import that fails. Set CIVITAI_API_KEY (or HF_TOKEN for Hugging Face) in ComfyUI's environment. Tokens are read from the environment only - never a widget, since widget values travel with every workflow you share.

Wallet and settings errors. PopNow 40201 means credits, not graph. PopNow 40004 means the model can't serve that combination and names the field. Both take ten seconds to read, and neither gets better by queueing again.