Qwen Image 2512 LoRA · Text to Image

Qwen-Image-2512 is the December 2025 refresh of Alibaba's 20B image model, and it's the one part of this family that gets underrated in public. The community's own line is that people argue endlessly about which model has the best realism while Qwen-Image 2512 and Wan 2.2 quietly stay their favourite for text-to-image and refining - with the well-known caveat that both are VRAM hogs and want beefier hardware. So a hosted node on 2512 is not a bad trade at all.

There's a second, more useful reason to pair 2512 with LoRAs specifically: this model responds unusually well to them. Not every base does. If you've trained or found a style LoRA, this is a good place to use it without buying a 24GB card first.

The inputs

• prompt - describe subject, composition and lighting. The tooltip is blunt about the trick: if a LoRA has a trigger word, put it first. Skipping that is the most common "my LoRA does nothing" report there is.
• loras (optional) - a SPICY_LORAS socket from the SpicyAPI LoRA node. Up to three weight files, each with its own strength, 0 to 4. The node chains, so you can stack a style and a character.
• aspect_ratio - default 1:1, nine shapes.
• output_format - jpeg, png or webp. Default is jpeg.
• seed - fixed after generate by default, so re-queueing the same graph reuses the first result instead of quietly charging you again.

There's no resolution tier on this one, unlike the 2.1 line - the model works at up to 1536 pixels on the long side, so shape is your only framing control. And no background toggle either, so don't go hunting for transparent output here; if that's what you need, use the 2.1 text-to-image node with png or webp.

Outputs

images and task id. Standard IMAGE value, straight into Save Image.

Wiring the LoRA

Add a SpicyAPI LoRA node and give it a direct https:// link to a .safetensors that opens without a login, plus a scale. Wire its output into loras here. Hugging Face resolve links work; a model page, a Drive folder, or a gated repo doesn't. Nothing reviews the weights, which is convenient and worth remembering - you're pointing a hosted service at an arbitrary file.

Install

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

Restart, then drop a key from spicyapi.ai/console/keys into Settings > SpicyAPI and restart once more. No dependencies to install, no weights to download, no GPU needed - the pack's dependency list is empty because it only uses what ComfyUI ships. Server installs should use SPICY_API_KEY in the environment rather than the settings panel.

Watch-outs

Strength: start at 1. The slider goes to 4 and the results at 4 are usually overcooked. Subtle style LoRAs often land best around 0.7; a character LoRA might want 1.2. Change one thing at a time.

Per-image billing at about $0.03 means batches are the risk. It's cheap enough to feel free and expensive enough that a hundred images isn't. Set Max cost per run under Settings > SpicyAPI - every run is quoted before submission, and anything over your ceiling stops with nothing charged.

Failure modes are pack-wide. No SpicyAPI key is set means add the key; 40310 means the account's email isn't verified; 40004 means no deployment serves the combo you picked. Keep the task id output when something fails - it's the exact handle for support. And interrupting a run in ComfyUI doesn't kill the job; it finishes server-side and is charged if it succeeds.