FLUX.1 [dev] SRPO

FLUX.1 [dev] SRPO is the community-loved "no plastic look" FLUX. It's a fine-tune of FLUX.1 dev post-trained with SRPO - Self-Play Reward Optimization, the RL-from-AI-feedback technique that judges its own outputs and reinforces what looks like a real photograph. That's the whole story of why it caught on: people kept describing it as realism without the waxy, airbrushed quality base dev is known for. It's hosted on Runware as runware:111@1, so you get it as a typed node without downloading the 12B stack or the fp8 drama that comes with it.

The honest caveat, before you're disappointed: it's still FLUX.1 dev under the hood, and 2026 has moved on - Klein and the Qwen-Image generation outclass it on complex scenes, and its day in the sun (late 2025) has passed. But for fast, genuinely photographic single-subject work, this remains a great API node. And the licensing reality of dev (non-commercial weights) becomes Runware's problem, not yours.

What you set

• positivePrompt (required) - full sentences work best; FLUX's T5 encoder eats natural language, not tag soup.
• width / height - 128–2048, step 64, default 1024². FLUX is happiest around its native square.
• steps / CFGScale - both gated toggles, off by default, letting the model pick. When you turn them on you're saying you know better; community lore says Euler-style schedulers with modest steps.
• seedImage - for img2img; wire an IMAGE in and strength controls how much the original survives.
• maskImage - with maskMargin (32–128 context pixels), this is your inpainting path.
• scheduler - 76 options, (default) first. Don't go exploring until you have a reason.
• lora, controlNet, ipAdapters, pulid, ultralytics sockets - the builder chain plugs in here for the full stacked-control workflow.
• layerDiffuse - transparent PNG output, if you're compositing.

Output is image (IMAGE), up to 20 per run via numberResults. advanced_json is the escape hatch for hiresFix.

How it works

Standard pack machinery: taskType: imageInference request over REST through the Runware SDK, references and masks encoded into the payload, native IMAGE back. The model AIR points at Runware's hosted SRPO checkpoint, so all the reward-optimized weight stays server-side. Cost prints on the title bar - SRPO was a beloved midweight, and it's priced accordingly.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

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

No model downloads; deps are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Blurry or muddy output on this node means you've gone off the default scheduler/CFG path - reset the gates to off and let the model's own plan run. Negative prompts do nothing here unless you force CFGScale on, because dev-style guidance is a different animal than SD-era CFG. And remember the licensing note: dev derivatives are non-commercial by the weights' license; if you need commercial, you want a different model AIR entirely.