A Geometric "Vector Network Analyzer" for Stable Diffusion Latent Trajectories.
This custom node implements the Manifold Alignment Protocol (MAP) within ComfyUI. It transforms the "black box" of diffusion sampling into a measurable, visualizable geometric process.
Instead of guessing whether Step=40 is better than Step=30, or which Scheduler works best, MAP-Probe quantifies the Semantic Confidence (Depth) and Structural Stability (Convergence) of the generation process.
For the theoretical foundation behind MAP, please visit the main repository:
π https://github.com/JBKing514/map_blog
This node requires matplotlib to generate visualization plots.
You MUST install it in the Python environment used by ComfyUI.
Open a terminal in your ComfyUI_windows_portable folder and run:
.\python_embeded\python.exe -m pip install matplotlibActivate your environment and run:
pip install matplotlibActs like an oscilloscope for your diffusion process.
- Differential Tracking: Automatically detects if you are using the same seed as the previous run.
- Visual Feedback: Displays Green (Improvement) or Red (Regression) to show exactly how your prompt or parameter tweaks affect generation quality.
- Quantified Metrics:
- Depth: How deep the latent vector penetrated the semantic manifold (Signal Strength).
- Stability: Whether the trajectory stabilized at the attractor well (Convergence Quality).
A "Self-Driving" mode that performs a 3-Phase Optimization:
- Step Search (Hill Climb): Iterates steps (e.g., 20, 25, 30...) to find the rough peak Q-score.
- Scheduler Sweep (New!): Once the optimal step count is found, it automatically tests all available schedulers (Normal, Karras, Exponential, SGM, etc.) to find the mathematically optimal solver for your model.
- CFG Refinement: Fine-tunes CFG scale around the optimal point with a user-definable range.
- Split-View Layout: Results are displayed in a professional, non-overlapping Dashboard below the trajectory plot.
- Phase Markers: Distinct visual markers for Steps (Line), Scheduler attempts (Triangles), and CFG trials (Stars).
- Auto-CSV: Successfully optimized runs are logged to
ComfyUI/output/MAP_Tuning_Log.csv.
Search for the node "MAP Pro Suite". It functions as a replacement for the standard KSampler.
- Input: Connect Model, Positive/Negative Conditioning, and Empty Latent.
- Output:
LATENT: The best latent found (passed to VAE Decode).IMAGE: The trajectory analysis plot (connect toPreview Image).STRING: Detailed analysis text.
Best for exploration and manual tweaking.
- Run with a Random Seed to establish a reference.
- Switch to Fixed Seed.
- Tweak parameters. The plot will show a Ξ (Delta) indicating improvement.
Best for finding the "Sweet Spot" of a Model/Prompt combo.
Parameters:
tuner_max_steps: Hard limit for step search (e.g., 50).tuner_stride: Step increment (e.g., 5).tuner_optimize_scheduler(New): If Enabled, the node will lock the best step count and brute-force test all schedulers to find the best ODE solver.tuner_cfg_range(New): Defines the search radius for CFG (e.g., 0.5 means checkingBase +/- 0.5).
Output: The node returns the Best Latent found across all three optimization phases.
MAP-Probe optimizes for Geometric Stability (Q-Score) β not beauty.
MAP focuses on improving the mathematical convergence quality of the sampling process:
- Clearer and more stable line art
- Smoother and more consistent shadows & gradients
- Reduced over-baking / ringing artifacts
- Better structural consistency across seeds
This does not always look βmore beautiful,β because:
- Some artistic styles (sketch / watercolor / noisy anime shading) naturally have low stability
- Human preference is nonlinear and style-dependent
- MAP rewards geometric clarity, not subjective appeal
Images with higher MAP-Q scores tend to:
- Convert much better in ESRGAN / SwinIR upscaling
- Handle hi-res fix with lower denoise (0.10β0.20) without collapsing
- Produce cleaner feature boundaries for downstream editing
- Enable more predictable prompt adjustments (good for workflows)
MAP acts as a navigator, not a dictator.
Use it to understand structure, then decide based on your own taste.
Below is a real-world comparison illustrating how MAP tuning affects clarity
(line sharpness, edge coherence, shadow smoothness), without dictating aesthetics.
Left: Default sampling (20 steps, 8 CFG, simple scheduler)
Middle: MAP-optimized sampling (25 steps, 8 CFG, exponential scheduler)
Right: Over-optimized sampling (60 steps, 12 CFG, simple scheduler)
MAP tuning generally improves edge definition without altering style.
Over-optimization introduces over-sharpening artifacts.
If you use MAP or its toolkits in your research, please cite:
@article{tang2025map,
title={Manifold Alignment Protocol (MAP) Specification},
author={Tang, Yunchong},
journal={Zenodo},
year={2025},
doi={10.5281/zenodo.18091447},
url={[https://doi.org/10.5281/zenodo.18091447](https://doi.org/10.5281/zenodo.18091447)}
}MIT License Β© Yunchong Tang