FLUX-FLUX Components Params

FLUX Defaults Hyper is a params node for the FLUX architecture, and it exists because FLUX isn't one model you blend - it's a bundle of four components: the diffusion model (here called diffuser), the clip_l text encoder, the t5xxl text encoder, and the vae. When you merge a FLUX checkpoint or bake a LoRA into one, each of those components is a separate key group, and they don't have to move at the same strength. This node hands you one recipe containing four independent weights, one per component.

The display name - "FLUX-FLUX Components Params" - looks like a typo but isn't: it's the auto-generated name for the flux-flux model config's components-params node, one of a whole family the pack generates for every supported architecture (there are SDXL Components Params and SD1 Components Params siblings).

What you set

Four FLOAT inputs, each a per-component strength (default 0.0, step 0.01):

• diffuser - the main FLUX transformer/diffusion model. The one you usually care about.
• clip_l - the CLIP-L text encoder half.
• t5xxl - the T5-XXL text encoder half.
• vae - the VAE.

Output is one recipe (MECHA_RECIPE) that you wire into a param socket - typically the alpha of a merge or LoRA-application node. The recipe is built by picking each component's value and OR-ing them together, so downstream, the merge applies diffuser's strength to diffuser keys, clip_l's strength to CLIP keys, and so on.

Why you'd bother

Most of the time one number for everything is fine - that's Float. The reason to split by component: LoRA merges and cross-model merges where the text encoders and the transformer respond differently. Baking a FLUX LoRA at alpha 0.8 can be great for the diffuser but oversaturated for T5; this lets you hold the transformer at 0.8 and drop the encoders to 0.5 without building a multi-branch workflow. It's also the natural target for "merge the transformer but not the VAE" workflows, since you can zero out components you don't want touched.

Install & gotchas

ComfyUI Manager → search mecha → Mecha Merge Node Pack, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart; only dependency is sd-mecha==1.1.7.

The trap: all four default to 0.0, and a components-params recipe with everything at zero is a "do nothing" recipe. If your FLUX merge silently doesn't change anything, check this node - you set diffuser and forgot the rest, or vice versa. Also note this node produces a param recipe; it goes into an alpha/param socket, not a weight socket. And while the pack knows the flux-flux key layout, it can only separate components if the models you're merging actually carry all four - a transformer-only FLUX file won't gain text-encoder keys just because you set a value for them.