Replicate black-forest-labs/flux-dev

Flux Dev is the middle tier of the Flux family and the one most people mean when they say "Flux" - strong prompt adherence, good detail, the default open-weights quality bar for a couple of years. This node runs it through Replicate's cloud instead of loading ~12GB of weights on your GPU, which is the whole appeal: full-fat Flux Dev on a laptop, nothing to download. It's a Replicate node, so you need a Replicate account and a REPLICATE_API_TOKEN, and you pay per generation.

It handles both text-to-image and image-to-image, so it's a genuine workhorse: better and more varied than Schnell, cheaper than Pro, and enough control to actually tune a result.

How it works

Dev is guidance-distilled from Pro. In practice that means there's a guidance scale but no negative prompt - the CFG-with-negatives mechanism was trained out. Pass an image and it switches to img2img, using prompt_strength to control how far it moves away from your source. Everything else is the familiar Flux control surface.

The inputs and outputs that matter

• prompt (required) - write natural language; Flux is at its best with full, descriptive sentences.
• aspect_ratio (1:1) - a ratio (16:9, 3:2, 9:16, …), not width/height fields.
• guidance (3) - prompt adherence. The default is on the low side deliberately; our Flux notes point out that the common 3.5 is often too high and drives the generic "Fluxmaxxed" faces, so 2–3 is frequently better.
• num_inference_steps (28) - quality vs cost. 28 is a solid default.
• image + prompt_strength (0.8) - supply an image for img2img; prompt_strength sets how much it changes.

Output is a single IMAGE (or up to 4 via num_outputs). go_fast (true) uses the optimized path, megapixels sets output size, output_format/output_quality set the file, disable_safety_checker drops the filter, seed fixes the result, and force_rerun bypasses Replicate's cache.

How to install it

One install brings every Replicate node.

• ComfyUI Manager: search ComfyUI-Replicate, install, restart.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/replicate/comfyui-replicate
  cd comfyui-replicate
  pip install -r requirements.txt
  

  then restart.

No Flux weights land on disk - the model lives on Replicate. Only the replicate pip package installs. Set your token first: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens).

Common issues & troubleshooting

Faces all look the same (and a bit alien). That's "Fluxmaxxing," and it's worst at high guidance. Drop guidance toward 1.8–2.5, cut beauty words ("beautiful," "gorgeous") from the prompt, and add real specifics about the person and scene. It's the single most effective fix and it costs nothing.

My negative prompt vanished. There isn't one - Dev is distilled and has no CFG negative. Say what you don't want by rephrasing the positive prompt.

img2img isn't changing / changing too much. That's prompt_strength. Higher moves further from your source image, lower keeps it. Pair it with a guidance you're happy with and iterate.

Blurry or muddy output. Usually too few steps for the scene, or guidance pushed too high. Nudge num_inference_steps up a little and guidance down. Same image on a re-run means the Replicate cache returned your earlier result - change the seed or set force_rerun.