ComfyUI Node
LTX Attn — Compare Runs
A ComfyUI node in g_raw/LTX/Profiler with 10 inputs and 2 outputs.
LTX Attn — Compare Runs
- diff_heatmap
- stats_text
◄store_handle_a►
◄store_handle_b►
◄attn_typesa►
◄metricentropy►
◄step_idx-1►
◄colormapdiverging►
◄cell_size16►
◄top_k15►
◄norm_percentile0.98►
◄diff_modeabsolute►
Categoryg_raw/LTX/Profiler
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| store_handle_a | STRING | — | |
| store_handle_b | STRING | — | |
| attn_type | COMBO | sa | 2 options: sa, ca |
| metric | COMBO | entropy | 12 options: entropy, temporal, spatial, sink, frame_dist_mean, frame_dist_std, +6 |
| step_idx | INT | -1-1–255 | — |
| colormap | COMBO | diverging | diverging: 0 = black, so identical cells read as neutral instead of coolwarm's near-white midpoint. |
| cell_size | INT | 164–64 | — |
| top_k | INT | 151–1536 | How many (block, head) pairs to list, ranked by the diff_mode score. |
| norm_percentile | FLOAT | 0.980.5–1 | Clip the diff_mode score beyond this percentile before mapping to color, so a few outlier cells don't wash out the rest of the heatmap to white. 1.0 = no clipping (use the true max). |
| diff_mode | COMBO | absolute | absolute: A - B, in the metric's own units. Not comparable across metrics with different intrinsic scales (e.g. sink in [0,1] vs raw temporal/spatial scores). relative_pct: (A-B) / max(|A|,|B|) * 100 -- % change, scale-free. zscore: (A-B) / std(A and B combined) -- diff in units of the metric's own spread, the most apples-to-apples way to ask whether one metric moved proportionally more than another. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| diff_heatmap | IMAGE | — |
| stats_text | STRING | — |