MiniMax H3 Image Edit

None of these MiniMax H3 nodes run a model on your machine. This one packages your prompt and your reference pictures, hands them to WaveSpeed AI over HTTPS, and drops the returned image back into the graph as a plain IMAGE tensor. No VRAM, no 42 GB of weights, no visible evidence that anything left the building. If that last part matters to you, stop here - because the whole mechanism is "your references go to a server you don't control."

What it buys you is the thing local workflows spent years bolting adapters together for: re-rendering a subject in a new scene, outfit or style while keeping the identity. IP-Adapter, FaceID, InstantID or a trained character LoRA were the old routes. The 2026 route is to paste references into an edit model's context and describe the change - and H3 Image Edit is the hosted version of that move. You cite each reference by number in the prompt (<Picture 1> through <Picture 9>), which is the convention the edit models actually read.

How it works under the hood

Wired images go through the pack's converter: each slice of the batch becomes a JPEG at quality 95, base64-encoded as a data: URI, which is what the API accepts anywhere it takes a file URL. images wins over image_urls if you wire both. The node then POSTs to /api/v3/wavespeed-ai/minimax-h3/image-edit, polls every 2 seconds, and gives up after 300 seconds - a five-minute ceiling, much shorter than the video nodes get. The result URL is decoded back into a tensor, and the node paints an inline preview on its own body.

Connect a MiniMax H3 LoRA Stack to the loras socket and the call reroutes to the endpoint's -lora twin instead, for an extra $0.015 an image.

The inputs you'll actually touch

prompt is required and is the only required field - that's the edit instruction, with the <Picture N> citations in it. images is a normal ComfyUI IMAGE batch (up to 9 references, first 9 win). image_urls is the URL fallback if your references already live online.

Then the three that decide what you get and what you pay: aspect_ratio (default auto, which follows your first reference; otherwise the same 15 ratios the Text-to-Image node offers), resolution (1k or 2k), and output_format (jpeg, png, webp). seed is sent to the API - -1 means a fresh one every queue. Output is a single image tensor; wire it to Save Image or into the next edit node.

What it costs

Roughly $0.03 at 1k and $0.09 at 2k, plus about $0.005 per extra reference beyond the first, plus $0.015 if a LoRA stack is attached. Two references at 2k with a LoRA is about $0.11 per click, so pin the seed once you like a result and let ComfyUI serve the cached output.

Installing it

ComfyUI Manager → Install Custom Nodes → search erpk → ERPK Custom Nodes → Install → restart. Or by hand:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

The pack is a monorepo - you're installing the WaveSpeed nodes plus the Claude, Gemini, OpenAI and Grok wrappers whether you want them or not, and the root requirements.txt pulls all five provider SDKs. The WaveSpeed side genuinely needs only requests, pydantic and Pillow, which is the set the pack's own troubleshooting section names.

Then add a WaveSpeed key. Right-click the canvas → ERPK Settings, or Settings → ERPK → API Keys. Keys live in your ComfyUI user settings, not in the workflow, so sharing the JSON doesn't leak them. If no key is configured anywhere, the client input is optional - but without a key you'll get Unauthorized: Invalid API key when you queue. You can also drop one in custom_nodes/erpk/wavespeed/config.ini under [API] WAVESPEED_API_KEY.

Nodes appear under ERPK → WaveSpeedAI.

Where people get burned

The 9-image cap is silent. Feed a 12-frame batch and references 10–12 just vanish. Same for a 20-URL image_urls string.

References are re-encoded to JPEG 95 on the way out. Fine for photos, visibly lossy for flat colour, line art, or anything with alpha - transparency doesn't survive.

auto isn't always what you meant. It follows the first reference, so a portrait reference gives you a portrait render even if the prompt describes a wide scene. Set the ratio explicitly when composition matters.

No prompt, no reference, no call. The node raises Prompt is required, then At least one reference image is required, before it ever touches the network.

It's a metered node in a loop. A batch of 8 at 2k is not a free experiment. Draft at 1k, then re-run the winner at 2k.