Replicate flux 1.1 pro ultra

Replicate flux 1.1 pro ultra is a wrapper for black-forest-labs/flux-1.1-pro-ultra running on Replicate's servers - the API-only flagship of the Flux 1 generation. The KB is clear about what that means: FLUX.1 Pro and 1.1 Pro were never released as weights; they exist only as BFL's API products. This node is the ComfyUI door into the top commercial tier, for people who want the flagship quality without signing up for BFL's own billing.

The mechanism is straightforward and notably well-built for a wrapper: the node fires off the Replicate call asynchronously - each image generation runs in a worker thread so the UI doesn't freeze - and supports batches, gathering multiple generations with asyncio.gather and collecting failures separately rather than letting one bad job kill the run. It also handles the awkward part of image-to-image over Replicate: the optional image_prompt tensor gets written to a temp file and passed as a file object, because that's what Replicate's API wants.

The inputs that matter:

• api_token - Replicate token (node input or REPLICATE_API_TOKEN env var).
• model - a dropdown with one choice, black-forest-labs/flux-1.1-pro-ultra. It's there because the source comments say you can add more Replicate models - the node is structured as a generic async Replicate runner with this model preloaded.
• prompt (default "A blackhole in space") - the generator.
• number_of_images (1–10) - batch size. This is the headline feature here: up to ten images per run, generated concurrently. The KB's caution about Flux batch behavior - quality falls off as you push batch counts up - applies, so treat 10 as "many tries in one bill," not a quality setting.
• image_prompt_strength (default 0.1, min 0.1) - how hard the optional reference image bites. This is the img2img dial, and the low default is worth internalizing: at 0.1 you're getting guidance from the reference, not a faithful copy.
• image_prompt (optional) - your reference image.
• seed (default -1 = random) and timeout (default 300s, up to 1800) - how long the node will wait on a run.

Outputs: image (the generated image, or the first one if batched - the source returns combined results), generation_info (JSON with the Replicate output and metadata). Files save to output/API/Replicate/.

Context worth having: Flux 1.1 Pro is the commercial half of the family whose open half (Dev, Schnell) reshaped local generation - the KB maps the whole lineage, including how Dev's guidance-distillation made it the default open checkpoint. The "ultra" tier was the flagship: the strongest text adherence and composition of its generation, with all the "Pro" caveats - closed, paid, no weights, and (per the KB's licensing coverage) BFL's increasingly restrictive terms hanging over the whole Flux line. If you want the flagship's quality without the model's baggage, this is the way.

Honest notes: it's paid per image, and batch runs are the fastest way to spend Replicate credits in this pack - a 10-image run is ten charges. If you're generating a bunch for iteration, consider whether one strong prompt beats ten weak ones. And because it's an API, network and queue time are part of the deal; the timeout input exists because ultra runs at high resolution can genuinely take minutes.

Install is the pack standard: Manager search "Node Fun", or clone https://github.com/nofunstudio/Node_Fun_ComfyUI into custom_nodes, pip install -r requirements.txt (the replicate dependency), restart. For the price of a token you get BFL's flagship in your graph - the API convenience the open-weights world can't give you.