🔧 Flux Model Blocks Buster

Flux's diffusion transformer isn't one monolithic unit - it's a stack of 19 "double" blocks (where the image and text streams still cross-attend separately) followed by 38 "single" blocks (where the two streams have merged into one). Different blocks in that stack tend to carry different jobs, and the LoRA block-weight crowd has spent a couple years mapping this out by trial and error on Flux LoRAs - one widely-repeated finding from that community work is that face and identity information tends to concentrate in a specific mid-range of the double blocks, while the very first and last blocks often carry comparatively little that's model-specific. Flux Model Blocks Buster takes that same per-block scaling trick and applies it directly to the base checkpoint's own weights, not just a LoRA sitting on top.

How it works

blocks is a big multiline text field, one line per block: ## 0 = 1.0 through ## 18 = 1.0 for the 19 double blocks, then # 0 = 1.0 through # 37 = 1.0 for the 38 single blocks. Each number is a scale multiplier on that block's contribution - 1.0 is identity (no change), push toward 0 to weaken or mute a block, above 1.0 to amplify it. The node ships with every block defaulted to 1.0, so you edit only the lines you actually want to change and leave the rest alone.

This is deliberately a probing tool, not a preset system - there's no button that says "isolate faces." You're expected to try ranges, look at what changed, and iterate, the same way the LoRA block-weight community arrived at their own findings.

Inputs and outputs

model and blocks. Outputs: MODEL (the patched checkpoint - wire it into your sampler like normal) and patched_blocks (a STRING confirming exactly which lines it recognized and applied). Check that second output after a run - it's your proof that your edits actually parsed, rather than silently doing nothing.

Installing it

Via ComfyUI Manager, search "ComfyUI Essentials". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cubiq/ComfyUI_essentials

Restart. Maintenance-only pack since April 2025 - cubiq isn't doing active development on it, so treat this as a stable but frozen tool.

Common issues & troubleshooting

Not sure where to even start. There's no settled recipe for this exact node, but the general community finding from LoRA block-weight experiments on Flux - a different tool, same underlying block structure - is a reasonable jumping-off point: the middle stretch of the double blocks tends to matter more for identity/subject content than the very first or last ones. Treat that as a hypothesis to test on your own images, not a rule.

patched_blocks came back shorter than the number of lines you edited. A line in your blocks text probably didn't parse - check for a typo in the ##/# prefix or the block index. A malformed line doesn't error, it just silently stays at its default weight, which is easy to miss if you don't check the output.

Results are subtle or you can't tell anything changed. Push the multiplier further from 1.0 before concluding a block doesn't matter - small nudges near 1.0 are genuinely hard to see in a single generation. Also generate at a fixed seed while you iterate, or you're comparing block changes against seed noise instead of against each other.