Selective LoRA Loader (FLUX)

Flux LoRAs are big, and they don't spend their weight evenly across the model's 57 transformer blocks. This node is the drop-in LoRA loader that lets you turn individual blocks off - so a face LoRA stops also dragging your composition and lighting along with it.

What it is and why you'd reach for it

This is the V1, standalone selective loader for Flux - a straight swap for ComfyUI's standard LoRA loader, except every one of Flux's 19 double blocks and 38 single blocks gets its own toggle and strength slider. The node's own description spells out the standard workflow: run the pack's separate "LoRA Loader + Analyzer" first to see which blocks a specific LoRA actually uses, then come here and switch off the ones that aren't pulling weight. Double blocks (0-18) typically carry more impact than single blocks (0-37), which the node states directly, but "typically" is doing real work in that sentence - every LoRA is different, which is the whole reason the analyzer step exists.

It's one of 31 nodes in ShootTheSound/comfyUI-Realtime-Lora. If you want analysis and selective loading combined into a single node instead of two wired together, the pack also ships a V2 "FLUX Analyzer + Selective Loader" that does both at once - this V1 node is for when you want the analyzer's output feeding a loader you control separately, or you're following an older workflow that already splits the two.

How it works

Connect the analysis_json output from the LoRA Loader + Analyzer node to this loader's own analysis_json input, and the block checkboxes color themselves by impact - blue for low, red for very high - so you can see at a glance where a LoRA's face, style, or composition weight actually lives before touching anything. Disabled blocks stay visible but dimmed, so you don't lose track of what you turned off.

Inputs and outputs that matter

• model / clip - standard Flux pipeline inputs, both patched by this node exactly like a normal LoRA loader.
• lora_name, strength - the usual pick and global weight.
• preset - jump straight to useful groupings instead of clicking 57 checkboxes: Double Blocks Only, Single Blocks Only, High Impact Double, Face Focus, Face Aggressive, Style Only (No Face), or Custom once you're hand-tuning.
• double_0 through double_18, single_0 through single_37, plus other_weights - each with its own toggle and a -5 to 5 strength slider.
• lora_path_opt (optional) - feed a path directly (from the pack's own trainer nodes, for instance) and it overrides the lora_name dropdown, which is the intended train-then-selectively-load pattern.
• analysis_json (optional) - the wire that turns on impact-colored checkboxes.
• Outputs: model, clip (patched, wire to your sampler and text encoder as normal), info (a text summary of what was applied).

Installing it

ComfyUI Manager: search "Realtime LoRA Trainer." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ShootTheSound/comfyUI-Realtime-Lora

Restart ComfyUI. This node works immediately - no extra dependencies, no model download. Those only apply to the pack's trainer nodes.

Troubleshooting

Checkboxes aren't colored, they're all plain gray. That coloring only appears once analysis_json is actually connected from an analyzer node. Run the LoRA Loader + Analyzer on the same LoRA first, or switch to the combined V2 "FLUX Analyzer + Selective Loader" node, which does the analysis inline without needing a second node.

lora_name is empty. Only LoRAs already sitting in your loras folder show up there. If you're loading one you just trained with this pack, use lora_path_opt instead of the dropdown.

Turning off "low impact" single blocks did nothing visible. Flux's single blocks are where text and image tokens interact directly, so even a low weight-magnitude score there can matter more than the number alone suggests. Treat the analysis as a strong hint, not gospel - walk changes back gradually and check the result at each step rather than disabling a large batch at once.