Z-Image Analyzer + Selective Loader V2

What it is

This is the pack's V2 answer to needing both LoRA Loader + Analyzer and Selective LoRA Loader (Z-Image) wired together - it folds analysis and per-layer control into one node, then adds two things neither V1 node had: strength scheduling (vary the LoRA's influence across the denoising steps, not just a flat value) and a save-to-disk option that lets you bake your block selection into a new, smaller LoRA file. If you're doing serious LoRA surgery on Z-Image rather than a quick strength tweak, this is the node to reach for over its V1 predecessors.

It also passes positive/negative conditioning straight through, which is the tell that this node is meant to sit inline in your main sampling graph rather than off to the side - you wire your prompt conditioning through it the same way you'd wire model and clip.

How it works

Same block-impact scoring as LoRA Loader + Analyzer underneath - 0–100% per layer based on weight magnitude, with the same 30-layer map the author's own testing mapped out (identity concentrating in layers ~18–28, style earlier, plus an "other weights" bucket and four extra Z-Image-specific blocks: context_refiner, noise_refiner, final_layer, x_embedder). What's new is the scheduling: instead of one strength number, you can supply a step:strength sequence like 0:0,1:1 (linear fade-in across the generation) or 0:1,0.5:1,0.6:0,1:0 (full strength for the first half, then cut), picked from over 40 built-in presets covering fades, bell curves, and pulses, or typed by hand.

Inputs and outputs that matter

• model, positive, negative, lora_name - this node sits inline with your conditioning, not just your model/clip.
• strength - the base multiplier (-10 to 10), before any schedule is applied.
• preset - the block-selection presets: Default, All Off, Half Strength, Late Only (20–29), Peak Impact (18–25), Face Priority, and more.
• layer_0–layer_29 (each with a _str slider), plus context_refiner, noise_refiner, final_layer, x_embedder, and other_weights - every one individually toggleable and reweightable.
• schedule_preset and strength_schedule (optional) - pick a named schedule or write your own step:strength sequence to vary influence during generation.
• save_refined_lora, save_path, save_filename (optional) - export your block-filtered result as a standalone LoRA file.
• Optional lora_path_opt overrides the dropdown with a direct path.

Outputs: model, positive, negative (all patched/pass-through, wire straight into your sampler), analysis (readable report), analysis_json (chain into another node if needed).

Installing it

Works out of the box with the rest of the pack - no external backend required. ComfyUI Manager (search "Realtime LoRA Trainer") or:

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

Restart ComfyUI. The pack's workflows/Lora Analysis, Block and Model Editing/ folder has a dedicated "ADVANCED" demo workflow built around this exact node's scheduling.

Common issues

Scheduling syntax silently does nothing. The format is strict: comma-separated step:strength pairs, steps expressed as 0–1 fractions of the total generation, not step counts. 0:0, 1:1 is a valid linear fade-in; a typo or an out-of-range step number tends to fail quietly rather than error loudly - start from a named preset and modify it rather than writing the string from scratch.

Extra Z-Image blocks (context_refiner, noise_refiner, etc.) get overlooked. These four sit outside the numbered layer_0–29 range and don't show up in most presets by default - worth checking individually if a LoRA still bleeds unwanted style after you've tuned the numbered layers.

Forgot to wire positive/negative through. Because this node passes conditioning inline, a broken conditioning chain here breaks your whole sampling graph, not just the LoRA - if generation looks wrong in an unrelated way, check this node is actually in the conditioning path, not just the model path.

save_refined_lora produces an unexpectedly large or small file. The saved LoRA only contains the blocks you left enabled - a heavily pruned selection produces a smaller file than the source LoRA, which is expected, not a bug.