ComfyUI Node
Enhanced KSampler (TJ)
A ComfyUI node in ✨ TJ_Node/Sampling with 21 inputs and 2 outputs.
Enhanced KSampler (TJ)
- model
- positive
- negative
- latent_image
- latent
- enhance_info
◄seed0►
◄steps20►
◄cfg8.0►
◄sampler_name▾►
◄scheduler▾►
◄denoise1.00►
◄enhance_enabledtrue►
◄enhance_archkrea2►
◄enhance_strength1.00►
◄enhance_debugfalse►
◄adv_text_scale1.00►
◄adv_active_scale1.00►
◄adv_per_token_whiten0.00►
◄adv_norm_equalize0.00►
◄adv_early_layer_scale1.00►
◄adv_mid_layer_scale1.00►
◄adv_late_layer_scale1.00►
Category ✨ TJ_Node/Sampling
Inputs (21)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| seed | INT | 00–18446744073709550000 | — |
| steps | INT | 201–10000 | — |
| cfg | FLOAT | 8.00–100 | — |
| sampler_name | COMBO | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 | |
| scheduler | COMBO | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 | |
| positive | CONDITIONING | — | |
| negative | CONDITIONING | — | |
| latent_image | LATENT | — | |
| denoise | FLOAT | 1.000–1 | — |
| enhance_enabled | BOOLEAN | true | Master switch. When off, every enhance option below is hidden and this behaves exactly like a standard KSampler. |
| enhance_arch | COMBO | krea2 | Which architecture's enhancer to use. Must match the loaded model — Krea2 patches the model's text-fusion adapter, while Klein and Z-Image operate on the positive conditioning. A mismatch is detected and skipped. |
| enhance_strength | FLOAT | 1.000–2 | Overall enhancement amount. 0 = no effect, 1.0 = default, 2.0 = maximum. Drives the built-in per-architecture profile; any Advanced knob you change below overrides that profile for this run. |
| enhance_debug | BOOLEAN | false | Print detection, applied architecture and tensor deltas to the console. |
| adv_text_scaleopt | FLOAT | 1.000.25–4 | [Krea2 only] Scale the whole text-MLP output, independently of Enhance strength (which only re-weights specific text chunks). 1.0 = unchanged. Raise it to push overall text influence up, lower it to let the base model speak more. Works even at strength 0. |
| adv_active_scaleopt | FLOAT | 1.000–10 | [Klein / Z-Image] Multiplier on every active-token embedding. 1.0 = unchanged. The model was trained on the text encoder's natural distribution, so values far from 1.0 push it off-distribution. |
| adv_per_token_whitenopt | FLOAT | 0.00-1–5 | [Klein / Z-Image] Amplify each token's deviation from the sequence mean: (x - mean) * (1 + w) + mean. Above 0 widens the spread (more contrast between tokens), below 0 compresses it. 0 = off. |
| adv_norm_equalizeopt | FLOAT | 0.000–1 | [Klein / Z-Image] Blend each token toward the sequence's mean L2 norm. Flattens magnitude variance, countering the text encoder's natural emphasis. 0 = off, 1.0 = fully equalized. |
| adv_early_layer_scaleopt | FLOAT | 1.000–5 | [Klein only] Scale the first Qwen3 layer slice (low-level / structural features). Klein conditioning stacks 3 layers along the embed dim. |
| adv_mid_layer_scaleopt | FLOAT | 1.000–5 | [Klein only] Scale the middle Qwen3 layer slice (intermediate semantic features). |
| adv_late_layer_scaleopt | FLOAT | 1.000–5 | [Klein only] Scale the last Qwen3 layer slice (high-level semantics). Raising this is usually what strengthens prompt adherence. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | — |
| enhance_info | STRING | — |