FLUX.1 Dev LoRA · Text to Image

FLUX.1 [dev] is the 12B model that replaced SDXL as the quality bar in 2024, and by 2026 it's the thing newer models get measured against rather than the thing you reach for first. Its remaining asset is the LoRA library - thousands of adapters trained against Dev, more than any of its successors have accumulated. This node runs Dev as a hosted text-to-image endpoint with up to three of your own LoRAs attached, $0.018 an image, about a megapixel in nine aspect ratios.

The case for it in 2026

Local Dev means a 12B stack, a text encoder, and the usual fight to fit it in VRAM at fp8 or GGUF - plus the fact that Dev's licence is non-commercial. If what you actually want is one specific DeviantArt-era LoRA that was never ported to anything newer, this node is a half-hour saved and no disk space spent. Paste the .safetensors URL into a SpicyAPI LoRA node, wire it in, and you're sampling Dev without owning Dev.

What you shouldn't do is use this as your default generator. Dev lost the photorealism lane to Z-Image, the complex-scene lane to Qwen-Image, and the editing lane to Klein and Qwen-Image-Edit. If you don't need a Dev-specific LoRA, one of those is probably the better spend.

The inputs

• prompt - subject, composition, lighting, in natural language. T5 is what made sentence prompting the norm for Flux, so write prose, not tag soup. Trigger word first if the LoRA has one.
• aspect_ratio - 1:1 by default, nine available; output lands around a megapixel.
• guidance_scale - 0 to 20, default 3.5. This is the canonical Flux Dev setting, so if you've tuned Dev locally, leave it alone.
• steps - 1 to 50, default 28. Same story as local Dev: fewer steps, faster and softer.
• output_format - jpeg (default), png or webp.
• seed.
• Optional loras - a SPICY_LORAS socket from a SpicyAPI LoRA node (URL + strength), chainable to stack, three max per request. Direct-download links only: if the URL needs a login, it won't load.

Outputs: images (IMAGE - save it or feed it downstream) and task id (STRING, the SpicyAPI task id, the thing support asks for).

Install

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

ComfyUI Manager → search SpicyAPI is the other route. Restart ComfyUI, then Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys, and restart once more so the node list is rebuilt from your account. Servers: SPICY_API_KEY. No weights, no dependencies - the pack's dependency list is empty and everything it needs, ComfyUI already has. Python 3.10+ and the V3 node API (tested with ComfyUI 0.39).

Two things to know about Dev specifically

First, the licence: FLUX.1 [dev] weights are non-commercial, and its T5 encoder is the famously filtered one - the reason the community's uncensored work moved to Chroma and Z-Image. A hosted endpoint doesn't change the licence that the weights carry, so do your own diligence if this is for paid work.

Second, LoRA stacking. Dev has a documented ceiling of roughly three subjects before a LoRA composition falls apart, and single-subject LoRAs trained on messy data are where it shines. Keep that in mind when you chain three adapters and wonder why the output is mush.

Spend-wise: set Max cost per run in the SpicyAPI settings, since $0.018 an image still adds up over a long sweep, and the pack quotes each request before submitting. Seeds start on fixed, so re-queueing the same graph reuses the cached image; randomize buys a new take. Failures look like No SpicyAPI key is set, 40310 (verify your email), 40004 (no deployment serves that combination), or a LoRA link that needs a login.