FreeU_LCM
Better detail from the fast model, one node
- pipe
- class
FreeU_LCM is a one-trick node, and the trick is a good one: it turns on FreeU inside the LCM pipeline. FreeU is a parameter-free inference technique that reweights the UNet's backbone and skip-connection features during denoising - it boosts detail and structure without retraining or extra models. Distilled models like LCM can look flat and oversmoothed in those 4 steps, so FreeU is a legitimately popular band-aid. This node is the pack's way of applying it to its own pipe.
How it works
It takes the pipe class output from any LCMLoader* node and four floats - the standard FreeU recipe:
- b1 (default 1.3) - backbone multiplier.
- b2 (default 1.4) - skip-connection multiplier (well, the pack maps b1/b2 into
enable_freeu, which is diffusers' standard b/s naming). - s1 (default 0.9) - backbone scaling factor.
- s2 (default 0.2) - skip scaling factor.
Under the hood it calls pipe.enable_freeu(s1=..., s2=..., b1=..., b2=...) and returns the same pipe, now FreeU-enabled. The defaults are the standard FreeU values for SD 1.5, which makes sense because LCM_Dreamshaper_v7 is a DreamShaper (SD 1.5-based) model. So you can honestly just plug it in, leave the defaults, and get the benefit.
Where it fits
Chain it right after your loader and before your generation node:
LCMLoader → FreeU_LCM → LCMGenerate (or any LCMGenerate*)
That's the whole wiring. If you find the pack's 4-step output a bit soft or lacking in texture, inserting FreeU is the first thing to try - it's free (no extra model, no training) and it usually recovers a surprising amount of crispness.
Setup and gotchas
Standard pack install - ComfyUI Manager (search LCM_Inpaint-Outpaint_Comfy) or clone + pip install -r requirements.txt from custom_nodes. No extra models. The pipe input is a class connection, so it only works inside this pack's pipeline - don't expect it to accept a native ComfyUI MODEL object.
Two honest notes. First, FreeU interacts with the oversaturation problem: if you're also running the pack's hot 8.0 CFG default, adding FreeU can push things further toward contrast-y mush. Drop CFG toward 1–2 and FreeU's detail boost becomes a clean win instead. Second, enable_freeu is a diffusers method, so it's tied to the diffusers-based pipelines in this pack - it won't apply to the SD-Turbo path the same way, and it's yet another sign this is 2023-era code doing its own thing outside native ComfyUI. For the pack's LCM nodes, though, it's a genuinely useful free upgrade.
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| b1 | FLOAT | 1.30–3 | — |
| b2 | FLOAT | 1.40–3 | — |
| s1 | FLOAT | 0.90–3 | — |
| s2 | FLOAT | 0.20–3 | — |
| pipe | class | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| class | class | — |