| # ComfyUI-H3NegativeLab |
|
|
| Negative prompting for MiniMax H3 in ComfyUI. |
|
|
| H3 is CFG-distilled — there's no stock negative prompt input. This pack patches the H3 DiT blocks to run a shadow attention branch with your negative text while keeping the positive presentation, references, and keyframes intact. NAG-form normalized repulsion is applied to the generated video and audio targets independently. |
|
|
| Works with T2V, FL2VA (first/last-frame I2V), and REF2VA. |
|
|
| ## Install |
|
|
| Extract in comfy\custom_nodes folder / Restart ComfyUI / No additional dependencies. |
| |
| ``` |
| ComfyUI/custom_nodes/ComfyUI-H3NegativeLab/ |
| __init__.py |
| h3_negative_lab_v1_0_2.py |
| README.md |
| ``` |
| |
| ## Nodes |
|
|
| ### H3 Shadow Negative |
|
|
| The everyday node. Plug in one negative CONDITIONING (from a plain CLIP Text Encode on the H3 text encoder), connect the patched MODEL to a normal BasicGuider, done. |
|
|
| `video_scale` and `audio_scale` control how hard the output is pushed away from the negative. Scale 1 = off for that stream. If `audio_scale > 1` and no separate `audio_negative` is connected, the main negative is reused for audio. |
|
|
| `tau` caps the guided norm relative to the positive (prevents overcooking). `alpha` blends between the positive-only and guided result. `start_percent` / `end_percent` limits guidance to a window of the denoise schedule. |
|
|
| ### H3 Shadow Negative (Advanced) |
|
|
| Same mechanism, explicit `video_negative` input. Optionally connect a separate `audio_negative`. |
|
|
| For REF2VA, two modes: |
|
|
| **Plain mode** — connect an ordinary H3 text CONDITIONING to `video_negative`. The patch preserves all packed REF2VA reference rows and replaces the user-prompt text in the shadow branch best-effort. |
|
|
| **Exact mode** — duplicate your stock `MiniMax H3 Reference to Video` node with the same refs, same sizing, same reference order, but put your negative description in its prompt. Connect that CONDITIONING to `video_negative`. The patch detects matching reference metadata and builds a full alternate packed layout with correct text/vision tags and RoPE coordinates for the shadow branch. |
|
|
| ### H3 Weighted Text Encode |
|
|
| `(phrase:weight)` syntax for H3 text conditioning. Operates in embedding space — weights above 1 emphasize, below 1 de-emphasize. The `(syntax:1.5)` is stripped before Qwen encoding. |
|
|
| Designed to feed the Shadow node's negative inputs. Safe alongside the Shadow block patch (it modifies the conditioning tensor, not the DiT blocks). |
|
|
| ## Usage |
|
|
| Minimal setup — add a negative to any H3 T2V workflow: |
|
|
| ``` |
| H3 CLIP Text Encode ("blurry, static, grainy") |
| | |
| v |
| H3 Shadow Negative <-- MODEL (after sigma shift if used) |
| | |
| v |
| BasicGuider <-- positive CONDITIONING |
| | |
| v |
| KSampler |
| ``` |
|
|
| Console output reports per-step metrics (neg-delta, guided-delta, clamp %) so you can tell whether your negative is actually doing anything. |
|
|
| ## Defaults |
|
|
| These worked well in testing as starting points: |
|
|
| | Parameter | Basic | Advanced | |
| |-----------|-------|----------| |
| | video_scale | 3.0 | 3.0 | |
| | audio_scale | 1.0 | 1.0 | |
| | tau | 2.5 | 2.5 | |
| | alpha | 0.35 | 0.35 | |
| | start_percent | 0.0 | 0.0 | |
| | end_percent | 0.60 | 0.60 | |
|
|
| If your output looks overcooked, lower alpha first, then tau. If the negative isn't biting, raise video_scale. |
| |
| ## Known limitations |
| |
| - **Block cache nodes**: This patch replaces H3's DiT double_blocks via `set_model_patch_replace`. Block-level cache nodes (Sage+Spectrum, T8mars, etc.) that replace the same blocks will conflict — use one or the other per run. |
| - **Experimental**: Keep fixed-seed A/B comparisons when tuning. The shadow branch adds one attention pass per active block per step; expect roughly 15-30% longer inference depending on your setup. |
|
|
| ## License |
|
|
| MIT |
|
|
