Bria Text2Image Fast

Nothing local runs when you use this node. Bria Text2Image Fast is a thin HTTP client dressed as a ComfyUI node: you hand it a prompt, it POSTs to Bria's hosted text-to-image/fast/2.3 endpoint, and it pulls back a finished image tensor. No checkpoint load, no sampler settings, no VRAM spike - your GPU is doing literally nothing except decoding a JPEG. That's the whole pitch, and if that sounds like a weird thing to put in the middle of a local-generation tool, you're not wrong. ComfyUI has been leaning into exactly this: the v0.25 release notes list a whole family of "Partner/API nodes" - Bria, Krea, Runway, Gemini - for people who want a hosted model inside their graph.

Why reach for it? Three honest reasons. You're on hardware that can't run modern checkpoints comfortably. You need something legally safe for a client job - Bria's whole selling point is models trained on 100% licensed data, no copyrighted characters or trademarks, which corporate buyers actually care about. Or you just want a second opinion without downloading another 12GB of weights. The name is not a lie about the mechanism - "Fast" is Bria's low-step generation tier, not a claim about your connection.

What you actually set

Only one input is required: api_key. The rest are optional, and the defaults are sane. The two that matter most for a beginner:

• prompt - your text. Bria's V1 models respond best to plain descriptive prompts; you don't need the SD tag-stack style.
• aspect_ratio - one of the nine standard ratios (1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9), default 4:3.

Then the knobs worth knowing: seed (-1 = random), steps_num (default 8 - that's the "Fast" part, Bria's distilled low-step pipeline), prompt_enhancement (set to 1 and it rewrites your prompt internally), and content_moderation (a 1 here runs a safety filter on the result).

The interesting inputs are the guidance slots. There are two independent guidance_method_* pairs, each choosing from controlnet_canny, controlnet_depth, controlnet_recoloring, or controlnet_color_grid - these are Bria's hosted ControlNet-analogues. They only activate when you also plug an IMAGE into the matching guidance_method_*_image input, and you set the influence with *_scale. Then there's image_prompt_mode (regular vs style_only) with image_prompt_image and image_prompt_scale for style transfer off a reference. So you can fake an img2img/ControlNet workflow entirely server-side, with zero local preprocessing.

The single output is output_image, a standard IMAGE tensor - wire it straight into Save Image or a Preview node like any local generation.

Installing

This is the easy part of the whole package. ComfyUI Manager → search "BRIA API" → install, or:

cd path_to_comfyui/custom_nodes
git clone https://github.com/Bria-AI/ComfyUI-BRIA-API

Restart ComfyUI. There are no model downloads and no heavy dependencies - the pack is just requests plus the usual torch/PIL stack you already have. Then grab an API token at platform.bria.ai and paste it into the node.

Where people get burned

The big one: the api_key field ships prefilled with the literal string BRIA_API_TOKEN, and leaving it there raises a "Please insert a valid API key" error. It's a placeholder, not a default that works.

Second, this is a paid hosted API - every run costs you a fraction of a cent at Bria's published per-request rates. Treat it like the commercial service it is: iterate locally, don't let a queue loop run 400 generations. Also note it's a synchronous call, so a slow Bria endpoint means ComfyUI sits there waiting; there's no progress bar from the model, just the request.

Third, temper expectations. This is Bria's legacy V1 pipeline - the README itself steers new work toward the V2 "FIBO" nodes, and early community reports on the underlying SDXL-derived model were blunt about it needing upscaling to look good. "Fast" gets you a quick, clean, commercially-safe image, not a portfolio piece. If you want Bria's best, that's a different node in this same pack; if you want the most control, you want a local checkpoint.

That's the honest trade: speed and licensing in exchange for control and money. For a deadline-safe corporate asset, it's a genuinely handy node. For your personal art, keep the local stack.