LCS Color Batch

The fastest way to color-match a grid of variations

You want the same prompt in five colors - a product shot in red, green, blue, and two more - without queueing five separate jobs. LCS Color Batch lets you do it in one batch, because it applies a different target color to each batch item. It's the batch-focused sibling of LCS Color Intervene, in the same facok/ComfyUI-LCS pack.

What it does

Feed it your model and the LCS calibration data (from LCS Load Data), type comma-separated hex colors, and it installs a post-CFG hook that steers each batch item toward its own color during sampling. Nothing runs after the VAE decode - the shift happens mid-generation in latent space, so it looks like the model intended the color instead of like a filter painted over the result.

The inputs you'll actually set

• colors - comma-separated hex, one per batch item: #FF0000,#00FF00,#0000FF. This is the entire point of the node.
• strength (0–2) - intervention intensity; 1.0 is full.
• mode - interpolated (default), type_i, or type_ii. Interpolated blends the two and is the sane starting point; type_i is a direct shift in the 3D color space for strong global shifts.
• start_step / end_step - the step window (defaults 8–10 of 50, the paper's sweet spot).

The second output is batch_size, and that's the detail that makes this flow: wire it into EmptyLatentImage's batch_size so your latent batch matches your color count.

The trap: color count vs batch size

batch_size is just len(your color list). Change EmptyLatentImage's batch size to something else and extra batch items ride along unsteered, producing colors you didn't ask for. Keep them in sync - that's literally why the output wire exists.

Install & setup

cd ComfyUI/custom_nodes
git clone https://github.com/facok/ComfyUI-LCS.git

Dependencies are einops and safetensors, typically already installed. You also need an LCS Load Data node; its first calibration run per VAE takes a minute, then it's cached on disk.

Honest take

Genuinely handy for reference sheets and product grids, but narrow. If you batch by prompt variation or a LoRA stack instead, you don't need it. And batch rendering multiplies VRAM - five colors is five latents in memory at once, so on a 8 GB card this will bite before it helps. Still, for a "give me this scene in every color" workflow, nothing in core ComfyUI does it this cleanly.