PVL FLUX General (fal.ai)

Most fal-node wrappers expose ten sensible settings and let the API do the rest. PVL FLUX General is not that. It wraps fal's flux-general endpoint - the one that hands you the sampler internals - with a control panel for scheduler, beta schedule, shift parameters, and NAG guidance, plus optional reference-image conditioning and up to three LoRAs. If you've ever wanted to tune a Flux generation below the level of "change the CFG," this is the node where the pack stops being a wrapper and starts being a lab bench.

It's part of the ComfyUI Assistant Node pack from pvlprk, in the fal family. The idea: same family of models as the rest, but with the dials most wrappers hide.

How it works

Under the hood it's the same pattern - build a payload, POST to fal-ai/flux-general, download the image. The difference is what goes in the payload. Besides the standard prompt, steps, CFG, num_images, and size (256–1440, defaulting to the 1392×752 landscape pair), you get:

• scheduler - euler or dpmpp_2m.
• use_beta_schedule and base_shift / max_shift - the Flux-family training-time shift knobs. For most people these should stay at defaults (off, 0.5, 1.15). They change the noise schedule, and "changed the noise schedule" is rarely the fix you needed.
• NAG parameters - nag_scale, nag_tau, nag_alpha, nag_end. NAG (negative-aware guidance) is the flux-general-specific guidance variant; these four control its strength and shape. This is deep-end stuff. Start with nag_scale and leave the rest.
• negative_prompt - yes, a real negative prompt, which regular Flux endpoints often lack.
• reference_image with reference_strength / reference_start / reference_end - image-conditioned generation: send a photo and a prompt, and the output follows both, with the strength sliders controlling how hard the reference pulls. The image goes out as a base64 data URI.
• lora1_name … lora3_name with scales - fal-hosted LoRAs, same as the pack's other LoRA nodes.

The inputs that matter for a beginner

Honestly: prompt, width/height, and if you want to experiment, reference_image plus reference_strength. Everything else has a default that's fine to leave alone. The shift and NAG parameters are for when you've read up on what they do and have a specific hypothesis - they're not tuning knobs you're supposed to wiggle blindly.

Output: one IMAGE.

Installing it

Part of pvlprk/comfyui-pvl-api-nodes. Install via ComfyUI Manager (search "ComfyUI Assistant Node") or:

cd ComfyUI/custom_nodes
git clone https://github.com/pvlprk/comfyui-pvl-api-nodes

Set your fal.ai key as FAL_KEY (or FAL_API_KEY / FAL_CLIENT_KEY), restart, add credits.

Where people get burned

The trap here is over-tuning. The NAG and shift parameters interact - change three at once and you won't know which one moved the result. Change one, keep notes. And this node's enable_safety_checker defaults to off, unlike the plain Flux Dev node, so if you're getting results you didn't expect from fal's side, check that toggle. Standard FAL_KEY and blank-placeholder-image gotchas apply as always.