Qwen Analyzer + Selective Loader V2

If you've ever loaded a Qwen-Image LoRA and gotten 90% of what you wanted plus one thing you didn't - a face that's a little off, a background style that leaked in when you only wanted the subject - this node is how you go find out exactly which part of the LoRA is doing that, and turn just that part down.

What it is

Qwen-Image's DiT has 60 transformer blocks, and a LoRA doesn't touch them evenly. Some blocks carry almost all of a character's likeness; others are quietly nudging the whole image toward a color palette or a composition habit you never asked for. This node loads a Qwen LoRA, scores every one of those 60 blocks by how much weight the LoRA actually put into it (0-100%, based on weight magnitude), and gives you a toggle plus a strength slider for each one - live, in the same node, no separate analyzer step required. That's the "V2" part: earlier versions of this pack split analysis and selective loading into two nodes you had to wire together; this one does both.

It's part of a much bigger pack - ShootTheSound/comfyUI-Realtime-Lora, "Realtime LoRA Trainer," 31 nodes covering training, analysis, block editing, and model debiasing across eight architectures. The author is active enough on Reddit that bug reports from a Saturday post get fixed and pushed the same afternoon.

How it works

Blocks aren't labeled by function, but the pack groups Qwen's 60 into rough zones you can reason about: 0-14 early (structure gets set here), 15-29 early-mid, 30-44 mid-late, 45-59 late (fine detail territory). There's also an other_weights slider for anything the LoRA touches outside the numbered blocks - worth using. The pack author has posted community findings on Z-Image showing style weight concentrating almost entirely in that "other" bucket while character likeness sat in a narrow band of numbered layers; Qwen behaves the same way often enough that skipping that slider means you're not actually seeing the whole LoRA.

Connect the analysis_json output from a plain analyzer (or feed one lora_path in through the optional input) and the checkboxes color themselves by impact - blue for low, red for very high - so you can eyeball which blocks matter before you touch anything.

Inputs and outputs that matter

• model / positive / negative - the base pipeline; this node patches all three because strength scheduling hooks through conditioning.
• lora_name / strength - standard LoRA pick and global strength.
• preset - quick zone presets (Late Only 45-59, Skip Early, Face-style groupings, etc.) instead of clicking 60 toggles by hand; set to Custom once you start hand-tuning.
• block_0 through block_59, each with its own on/off and a -5 to 5 strength slider, plus other_weights.
• strength_schedule (optional) - a step:strength list like 0:0,1:1 to fade the whole LoRA in or out over the generation, with 40+ named presets to pick from instead of typing your own curve.
• save_refined_lora (optional) - export whatever block combination you landed on as its own .safetensors file, so you don't have to rebuild the same toggle state every session.
• Outputs: model, positive, negative (patched, wire straight into your sampler), analysis (readable text breakdown), analysis_json (for chaining to other selective loaders).

Installing it

Search "Realtime LoRA Trainer" in ComfyUI Manager, or the manual route:

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

Restart ComfyUI. This node needs nothing beyond that - no training backend, no extra Python packages. That requirement only kicks in for the pack's trainer nodes.

Troubleshooting

The lora_name dropdown is empty. It only lists files ComfyUI can already see in your loras folder - if you haven't dropped a Qwen-Image LoRA in there yet, or you're feeding a freshly trained one via the lora_path_opt string input instead, the dropdown having zero choices is expected, not broken.

Disabling blocks tanked the likeness instead of the style. A well-trained LoRA can genuinely spread its signal across most of the network; nuking half the blocks because they scored "medium" impact will cost you more than it fixes. Treat the color-coded impact as a starting hypothesis, not a rule - disable the lowest-scoring handful first, generate, and walk back up only if the problem persists.

Strength schedule doing nothing. The format is strict: step:strength pairs, comma-separated, steps expressed as a 0-1 fraction of the run - 0:0, 1:1 fades in, not 0,1:0,20 in raw step counts. Reach for the preset dropdown before hand-writing a curve; there are over 40 already built, including bell curves and step functions for exactly this.