How does ComfyUI implement the Conditioning (Set Mask) node?

[Pauan]

I've spent hours looking through the source code, but I can't find the actual implementation, maybe I'm blind and just missing it.

The node itself just sets the "mask" cond, but I can't find where the "mask" cond is being applied.

I'm trying to figure out if ComfyUI is using the same implementation as the webui regional prompter and if not, what the differences are.

[gershu-ar]

Correct me if I'm wrong: you’re not missing it — the mask node itself only sets the “mask” cond. The actual application happens inside the inpainting pipeline, specifically in the VAEEncodeForInpaint node. That’s where the image and mask are encoded together into latents, and the sampler then generates only within the masked regions.

So ComfyUI is not using the same implementation as WebUI’s regional prompter. Regional prompter works by splitting prompts across attention regions, while ComfyUI inpainting uses explicit spatial masks combined with inpainting-trained models.

In short: mask cond in ComfyUI = latent encoding with mask, not prompt-based conditioning.

[Pauan]

The actual application happens inside the inpainting pipeline, specifically in the VAEEncodeForInpaint node.

Hmmm, that doesn't sound right. Even if I don't use the VAEEncodeForInpaint node the Conditioning (Set Mask) node still works fine.

Also, VAEEncodeForInpaint doesn't actually do the implementation either, it only sets the "noise_mask", the implementation is somewhere else.

I'm trying to find where in the code it's actually processing things like "mask", "noise_mask", etc.

[vectornguyen76]

It's in comfy/samplers.py, not in any node. The node only stores mask, mask_strength and set_area_to_bounds on the cond, and the sampler uses the mask to weight the model's output per cond at every step.

1. When sampling starts, process_conds calls resolve_areas_and_cond_masks_multidim (here, here). It resizes each cond's mask bilinearly to the latent size, and with set_cond_area = mask bounds it turns the mask's bounding box into an area crop.
2. On each step, get_area_and_mult (L33-L79) builds mult = mask * mask_strength * strength for that cond, cropped to the area if there is one. Conds without a mask get torch.ones_like(x).
3. calc_cond_batch runs the model once per cond (batched when shapes allow), then does out += output * mult, count += mult, and at the end out /= count (L336-L354).

So for each latent pixel the positive prediction is sum(pred_i * mask_i) / sum(mask_i) over the positive conds, and the negative side is blended the same way before CFG. The prompts never meet inside attention. Each one gets its own full forward pass over the whole latent (or over the cropped box with "mask bounds"), and the masks only weight the outputs.

A side effect: a pixel that no mask covers ends up with output 0 (the count is initialised to 1e-37), so in practice you combine masked conds with an unmasked base prompt. Conds marked as default via Cond Set Default Combine (nodes_hooks.py) are used to fill exactly that leftover area (finalize_default_conds).

Compared with the regional prompter, this is not its Attention mode, which splits regions inside the cross-attention layers in a single pass. It's closer to its Latent mode, which also runs a separate pass per region (its README says generation time scales with the number of areas). I haven't compared the blending math between the two line by line.