Latent¶

Nodes that operate on LATENT tensors: value shaping, per-channel adjustments, and inspection.

At a glance¶

Node	Summary
Latent Clamp	Clamps every latent value to a min/max range to tame "burn" artifacts.
Latent Scale	Multiplies all latent values by a scalar to raise or lower overall energy.
Latent Normalize	Applies one of 15 normalization methods, blended by strength, whole-tensor or per-channel.
Latent Stats	Prints shape/min/max/mean/std to the console and passes the latent through.
Latent Channel Offset	Adds a per-channel bias to the first 4 latent channels.
Latent Channel Offset x16	The 16-channel version, for Wan/Flux/SD3-class latents.
Latent Channel Scale	Multiplies each of the first 4 channels by its own factor (negative inverts).
Latent Channel Scale x16	The 16-channel multiplicative version.

Value shaping¶

Latent Clamp¶

Hard-limits every value in a latent to a [min, max] window.

Diffusion latents can develop a few extreme outliers that decode as "burn" — blown-out highlights or oversaturated blotches. Clamping the tails back into a sane range tames those artifacts without reshaping the rest of the distribution. The before/after min–max is logged to the console.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input (cloned, never mutated).
`min_value`	FLOAT	-4.0	Lower bound (`-20`–`0`).
`max_value`	FLOAT	4.0	Upper bound (`0`–`20`).

Outputs: latent.

Latent Scale¶

Multiplies all latent values by a single scalar.

Amplifying (>1) or attenuating (<1) latent magnitude changes the overall contrast/energy of the decoded result — e.g. nudging latent energy before a second sampling pass. A scale of exactly 1.0 passes the input through untouched.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input.
`scale`	FLOAT	1.0	Multiplier (`0`–`10`).

Outputs: latent.

Latent Normalize¶

Rescales or re-centers a latent with one of 15 normalization methods, then blends the result back toward the original by strength.

Useful for balancing latent distributions, reducing artifacts, or matching statistics between stages. The method list is deliberately identical to Conditioning Normalizer for consistency across latent and conditioning workflows.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input.
`method`	choice	`none`	See methods below.
`strength`	FLOAT	1.0	Blend between original (`0`) and fully normalized (`1`). Values outside 0–1 extrapolate.
`per_channel`	BOOLEAN	false	Normalize each channel independently instead of using whole-tensor statistics.

Methods: max_norm (divide by max absolute value), std_norm (divide by standard deviation), std_half (gentler, divide by 2×std), zscore (subtract mean, divide by std), zscore_avg (average of z-score and max-norm), zscore_half (gentler z-score), slight_z (20% z-score, 80% max-norm), mean_restore (normalize but restore the original mean), range (rescale to [-1, 1]), and hard clamps clamp_1, clamp_1.5, clamp_2, clamp_3, clamp_4.

Outputs: latent.

Latent Stats¶

Prints a latent's shape, min, max, mean, and standard deviation to the console and passes it through completely unchanged.

Drop it inline anywhere in a graph to inspect values.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input (true pass-through).
`label`	STRING	`Latent`	Header printed above the stats, to tell multiple stat nodes apart.

Outputs: latent (unchanged).

Channel adjustment¶

Latent channels carry loosely semantic content. For SD 1.5/SDXL's 4-channel latents the rough correlations are: ch0 ≈ cyan–red / brightness-openness, ch1 ≈ magenta–green / structure, ch2 ≈ yellow–blue tones, ch3 ≈ luminance/contrast — but the effects are abstract and semantic, not literal color shifts. The x16 variants exist for 16-channel latents (Wan and other newer-generation VAEs), whose channel meanings are less documented — experimentation encouraged. All four nodes safely ignore channels the latent doesn't actually have.

Latent Channel Offset¶

Adds a per-channel constant bias to the first four latent channels.

Positive values push a channel toward one end of its spectrum, negative toward the other, shifting the abstract content that channel encodes. Only channels with a non-zero offset are touched.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input.
`channel_0`–`channel_3`	FLOAT	0.0	Amount added to each channel.

Outputs: latent.

Latent Channel Offset x16¶

The same additive offset with 16 channel controls (ch_00–ch_15), for 16-channel latents.

Lets you probe or adjust individual channels of the wider latent. Controls beyond the latent's real channel count are harmlessly ignored.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input.
`ch_00`–`ch_15`	FLOAT	0.0	Amount added to each channel.

Outputs: latent.

Latent Channel Scale¶

Multiplies each of the first four channels by its own factor.

Unlike offset (which shifts), scaling changes a channel's amplitude: >1 amplifies its contribution, <1 reduces it, and negative values invert its polarity. The default of 1.0 is a true no-op.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input.
`channel_0`–`channel_3`	FLOAT	1.0	Multiplier for each channel.

Outputs: latent.

Latent Channel Scale x16¶

The 16-channel multiplicative counterpart, for Wan-style latents.

Same amplify/reduce/invert semantics per channel.

Parameter	Type	Default	Description
`latent`	LATENT	—	Input.
`ch_00`–`ch_15`	FLOAT	1.0	Multiplier for each channel.

Outputs: latent.